Intelligent classification method and system for application materials based on data analysis

Abstract

The application discloses an application material intelligent classification method and system based on data analysis. The method comprises the following steps: after obtaining the application material text, firstly, extracting keywords and calculating the weight to generate a weighted keyword vector. The vector is mapped to a high-dimensional semantic space and converted into a multi-dimensional semantic representation vector. The semantic similarity sequence of the vector and the preset category standard vector is calculated, and preliminary classification is carried out by comparing with the high similarity threshold. If the matching is successful, the corresponding type is output as the final result. If the matching fails, the keyword weight is re-evaluated, and the enhanced semantic representation vector is generated. The similarity of the vector and the standard vector is calculated again, and the abnormal material is screened out according to the low threshold. Finally, the semantic proximity analysis and consistency evaluation grouping of the abnormal material are carried out, and the final classification is obtained. The method and system can effectively improve the classification accuracy of various application materials in the field of digital public service.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to an intelligent classification method and system for application materials based on data analysis. Background Technology

[0002] In fields such as digital public services, how to leverage big data analytics and other technologies to achieve highly consistent automatic classification of various application materials, thereby reducing the burden of manual review, has become a current research hotspot.

[0003] One existing intelligent classification method for application materials primarily involves the system pre-creating a keyword list for each material category. These keywords are typically set manually based on experience or extracted from a small sample. When processing materials, the system segments the submitted text and counts the frequency of each of these pre-defined keywords. Then, the system calculates a matching score between the text and each category, usually based on a simple weighted average or direct summation of keyword occurrences. Finally, the system assigns the materials to the category with the highest matching score, completing the automatic classification. This entire process relies heavily on a pre-defined vocabulary and has a relatively simple weighting process. Traditional techniques, which depend on fixed keywords and simple frequency statistics, struggle to handle the flexibility and variability of expressions, leading to semantic representation bias and fluctuations in similarity calculations. Therefore, this existing method cannot effectively improve the classification accuracy of application materials. Summary of the Invention

[0004] This invention provides a data analysis-based intelligent classification method and system for application materials to improve the accuracy of material classification.

[0005] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an intelligent classification method for application materials based on data analysis, comprising: Obtain the application materials text; Based on the application materials text, keyword extraction, weight calculation and weighting are performed to obtain a weighted keyword vector; Based on the weighted keyword vector, the weighted keyword vector is mapped to a high-dimensional semantic space by capturing the contextual associations between words to obtain a multi-dimensional semantic representation vector. Based on the multidimensional semantic representation vector and the preset semantic category standard vector, the semantic similarity between vector directions is calculated one by one to obtain a semantic similarity sequence; Based on the semantic similarity sequence and the preset semantic high similarity threshold, material type matching analysis is performed to determine the preliminary material classification result; When the preliminary classification result of the material shows that the type match is successful, the material type is output as the final classification result. When the initial classification result of the material shows that the type matching fails, the distinguishing weight of the keywords is re-evaluated according to the application material text, and the multidimensional semantic representation vector is supplemented and strengthened according to the re-evaluated distinguishing weight to form an enhanced semantic representation vector. Based on the enhanced semantic representation vector, a new semantic similarity sequence is obtained by calculating the semantic similarity between the enhanced semantic representation vector and the semantic category standard vector, and abnormal application materials are screened out in combination with a preset semantic low similarity threshold. Based on the abnormal application materials, semantic proximity analysis is performed, and the groups are grouped according to the semantic proximity analysis results through semantic consistency evaluation to obtain the final classification results.

[0006] Secondly, the present invention provides an intelligent classification system for application materials based on data analysis, comprising: The data acquisition module is used to obtain the text of application materials; The keyword analysis module is used to extract keywords, calculate and weight them according to the application material text, and obtain a weighted keyword vector. The semantic feature analysis module is used to map the weighted keyword vector to a high-dimensional semantic space by capturing the contextual associations between words, based on the weighted keyword vector, to obtain a multi-dimensional semantic representation vector. The semantic similarity analysis module is used to calculate the semantic similarity between vector directions one by one based on the multidimensional semantic representation vector and the preset semantic category standard vector, so as to obtain a semantic similarity sequence. The preliminary classification module is used to perform material type matching analysis based on the semantic similarity sequence and a preset semantic high similarity threshold to determine the preliminary material classification result; The preliminary output module is used to output the material type as the final classification result when the preliminary classification result of the material shows that the type match is successful. The enhanced semantic module is used to re-evaluate the distinguishing weight of keywords based on the application material text when the preliminary classification result of the material shows that the type matching fails, and to supplement and strengthen the multi-dimensional semantic representation vector based on the re-evaluated distinguishing weight to form an enhanced semantic representation vector. The anomaly screening module is used to obtain a new semantic similarity sequence by calculating the semantic similarity between the enhanced semantic representation vector and the semantic category standard vector based on the enhanced semantic representation vector, and to screen out abnormal application materials in combination with a preset semantic low similarity threshold. The final classification output module is used to perform semantic proximity analysis on the abnormal application materials, and group them according to the semantic proximity analysis results through semantic consistency evaluation to obtain the final classification results.

[0007] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention extracts keywords and integrates word frequency weight, inverse document frequency weight and statistical variance discrimination weight to construct a weighted keyword vector. It combines positional encoding and self-attention mechanism to generate a high-dimensional semantic representation, thereby improving the discriminative ability and semantic consistency of text feature expression.

[0008] (2) This invention adopts a high and low dual threshold matching mechanism and semantic bias analysis to achieve two-stage discrimination of preliminary classification and enhanced classification. Combined with context sliding window and feature matrix reconstruction, it strengthens semantic representation and improves the classification coverage and matching accuracy of flexible and semantically ambiguous materials.

[0009] (3) Based on DBSCAN density clustering, the present invention performs semantic proximity grouping of abnormal materials, and achieves automatic classification of unconventional and marginal case materials by evaluating the consistency within the group and matching the standard category with the overall semantic vector, thereby improving the system's adaptability to complex application materials and the completeness of classification.

[0010] (4) This invention constructs a complete semantic representation-similarity matching-abnormal clustering-result output process, integrating supervised similarity calculation and unsupervised clustering analysis, which improves the overall classification robustness and system practicality of diverse and dynamically changing application materials while ensuring high accuracy. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the intelligent classification process for application materials based on data analysis provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the intelligent classification structure for application materials based on data analysis provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides intelligent classification of application materials based on data analysis, including the following steps: S11, Obtain the application materials text; S12, Based on the application material text, perform keyword extraction, weight calculation and weighting to obtain a weighted keyword vector; S13, Based on the weighted keyword vector, the weighted keyword vector is mapped to a high-dimensional semantic space by capturing the contextual associations between words to obtain a multi-dimensional semantic representation vector; S14. Based on the multidimensional semantic representation vector and the preset semantic category standard vector, calculate the semantic similarity between vector directions one by one to obtain a semantic similarity sequence. S15, Based on the semantic similarity sequence and the preset semantic high similarity threshold, perform material type matching analysis to determine the preliminary material classification result; S16, when the preliminary classification result of the material shows that the type matching is successful, the material type is output as the final classification result; S17, when the preliminary classification result of the material shows that the type matching fails, the distinguishing weight of the keywords is re-evaluated according to the text of the application material, and the multidimensional semantic representation vector is supplemented and strengthened according to the re-evaluated distinguishing weight to form an enhanced semantic representation vector. S18. Based on the enhanced semantic representation vector, a new semantic similarity sequence is obtained by calculating the semantic similarity between the enhanced semantic representation vector and the semantic category standard vector, and abnormal application materials are screened out in combination with a preset semantic low similarity threshold. S19, Based on the abnormal application materials, semantic proximity analysis is performed, and the groups are grouped according to the semantic proximity analysis results through semantic consistency evaluation to obtain the final classification result.

[0014] In step S11, the application materials text is obtained.

[0015] Specifically, the acquisition of application material texts is achieved through data collection. A pre-set application material database is connected, storing application material records accumulated from the historical application system. Each record contains an application identifier and corresponding text content. When a classification task instruction is received, the corresponding application material text content is retrieved from the application material database based on the application identifier attached to the instruction. The application material text is a complete application content stored in string form, including structured or unstructured text information such as the application title, applicant information, reason for application, and attachment descriptions. This step is the starting point of the classification process, providing raw text input for subsequent keyword extraction and semantic analysis.

[0016] In step S12, based on the application material text, keyword extraction, weight calculation, and weighting are performed to obtain a weighted keyword vector, including: Based on the application materials text, text segmentation and vocabulary unit extraction were performed using Chinese word segmentation technology to obtain the word segmentation results; Based on the word segmentation results and the pre-stored corpus document frequency statistics, the product of the word frequency weight and the inverse document frequency weight of each word unit is calculated using the word frequency-inverse document frequency algorithm to generate a keyword weight sequence. Based on the keyword weight sequence, the statistical variance of the frequency of each word unit in the application material text is calculated to obtain the distinguishing weight of each word unit; The word frequency weight, the inverse document frequency weight, and the differentiation weight are weighted and fused to obtain the key feature vector; The key feature vectors that are greater than the preset key feature threshold are selected to form an optimized feature set; Based on the optimized feature set, normalization and vector dimension alignment are performed to obtain a weighted keyword vector.

[0017] Specifically, based on the application material text output in step S11, keyword analysis is performed, starting with Chinese word segmentation. A pre-built Chinese word segmentation dictionary is invoked; this dictionary is derived from a historically accumulated application material corpus, constructed through word frequency statistics and domain terminology extraction. The segmentation process employs a maximum forward matching algorithm based on this dictionary, scanning and matching the application material text character by character, dividing the continuous character sequence into independent lexical units, resulting in an ordered set containing all lexical units—the segmentation result.

[0018] Subsequently, preliminary weights are calculated based on the word segmentation results and pre-stored corpus document frequency statistics. The corpus document frequency statistics are obtained by analyzing a historical application material corpus, counting the number of documents in which each lexical unit appears in the entire corpus. The system calculates a word frequency weight for each lexical unit in the application material text; this weight equals the number of times the lexical unit appears in the current application material text divided by the total number of lexical units in the text. Simultaneously, the system calculates the inverse document frequency weight for each lexical unit; this weight equals the total number of documents in the corpus divided by the number of documents containing that lexical unit, and then taking the logarithm to base 10 of that quotient. The word frequency weight and the inverse document frequency weight of each lexical unit are multiplied to generate a keyword weight sequence containing each lexical unit and its corresponding product value.

[0019] The system further calculates the distinguishing weight of each lexical unit to assess its distribution specificity within the text. This process first counts the position index of each lexical unit in the application material text for each occurrence, and then calculates the statistical variance of these position index values. This variance reflects the dispersion of the lexical unit's distribution in the text; a larger variance indicates a more concentrated distribution and a greater potential for distinguishing specific topics. This variance is then linearly normalized to fall between 0 and 1, yielding the distinguishing weight of each lexical unit. Specifically, the application material text is divided into M text segments (such as sentences or paragraphs) based on punctuation marks (e.g., periods, semicolons) or a fixed length, and the occurrence frequency of the target lexical unit in each text segment is counted (which can be 0). Based on the sequence of occurrences of the word in these M segments, the variance of this sequence is calculated; a larger variance indicates a more concentrated distribution of the word in the text and a greater potential for distinguishing text topics or categories.

[0020] The system weights and fused weights based on word frequency, inverse document frequency, and discriminative weights to generate a comprehensive key feature vector. This fusion process uses three pre-defined sets of fusion weight factors, corresponding to word frequency, inverse document frequency, and discriminative weights, respectively. These fusion weight factors are determined through grid search and backpropagation, utilizing the impact of different weight combinations on classification accuracy in historical classification tasks. For each lexical unit, its word frequency weight is multiplied by the word frequency fusion weight factor, its inverse document frequency weight is multiplied by the inverse document frequency fusion weight factor, and its discriminative weight is multiplied by the discriminative fusion weight factor. The three products are then summed to obtain the key feature value for that lexical unit. All lexical units are arranged in lexicographical order, and their key feature values ​​constitute the key feature vector.

[0021] The system filters key feature vectors based on a preset key feature threshold to form an optimized feature set. The key feature threshold is set as the 75th percentile of all key feature vector values ​​in historical samples, sorted by size. This threshold is determined by statistically analyzing the distribution of all key feature values ​​generated during the processing of historical application materials. The system iterates through each key feature value in the key feature vectors, retaining only items greater than the threshold and their corresponding lexical unit indices, and discarding items less than or equal to the threshold. The retained items constitute the optimized feature set.

[0022] Finally, the system normalizes and aligns the vector dimensions of the optimized feature set to generate a weighted keyword vector. Normalization involves subtracting the minimum value of all key feature values ​​in the set from each key feature value in the optimized feature set, and then dividing by the difference between the maximum and minimum values, mapping all values ​​to the range of 0 to 1. Vector dimension alignment is achieved by querying the global vocabulary, which contains all word units that have appeared in the corpus and assigns them unique dimension indices. The system creates a zero vector with a length equal to the total number of words in the global vocabulary. Then, based on the index position of each word unit in the optimized feature set in the global vocabulary, the corresponding normalized key feature value is filled into the corresponding position of the zero vector, leaving the remaining positions as zero. The resulting fixed-dimensional, sparse numerical vector is the weighted keyword vector, and semantic feature analysis is performed on this vector in step S13.

[0023] In step S13, based on the weighted keyword vector, the weighted keyword vector is mapped to a high-dimensional semantic space by capturing the contextual relationships between words to obtain a multi-dimensional semantic representation vector, including: Based on the weighted keyword vector, calculate the sine and cosine position encoding vector for each word, and add the sine and cosine position encoding vector to the weighted keyword vector to obtain a position enhancement vector sequence; The position enhancement vector sequence is input into a pre-built deep semantic encoder based on a self-attention mechanism. The encoder performs nonlinear transformation and context encoding through a multi-layer self-attention network and a fully connected network to output the hidden layer state tensor. Based on the hidden layer state tensor, the association strength of lexical features in different semantic subspaces is calculated through a multi-head self-attention mechanism, and the position enhancement vector sequence is weighted and fused to obtain the context feature vector. Semantic discrete judgment is performed based on the context feature vector, and vector reconstruction is performed based on the semantic discrete judgment result to obtain a multidimensional semantic representation vector.

[0024] In one specific implementation, the step of performing semantic discrete judgment based on the context feature vector and reconstructing the vector based on the semantic discrete judgment result to obtain a multidimensional semantic representation vector includes: Calculate the numerical variance of the context feature vector in each dimension to obtain the context feature variance; When the variance of the context features is not greater than a preset semantic discrete threshold, the context feature vector is directly determined as a multidimensional semantic representation vector. When the variance of the context features is greater than the preset semantic discrete threshold, nonlinear dimensionality reduction and cluster separation are performed based on the context feature vector using an equal metric mapping algorithm, and the low-dimensional manifold vector obtained after dimensionality reduction is reconstructed to the original dimension to obtain a multidimensional semantic representation vector.

[0025] Specifically, semantic feature analysis calculates positional encoding vectors based on the input weighted keyword vectors. Each word vector in the weighted keyword vector used in this process has a dimension d_model of 512. The dimension of the positional encoding vectors is consistent with d_model. For the word with the sequential index pos in the segmentation results, the positional encoding vector is calculated in the i-th dimension. When i is even, a sine function value is calculated. When i is odd, calculate the cosine function value. The positional encoding vector of each word is element-wise added to its corresponding weighted keyword vector to generate a sequence of position-enhanced vectors. This operation incorporates the absolute positional information of words into the features, enabling the model to perceive the impact of word order on semantics.

[0026] It should be noted that the weighted keyword vector output in step S12 has an original dimension equal to the size of the global vocabulary. To adapt to the subsequent deep semantic encoder based on the self-attention mechanism, this high-dimensional sparse vector needs to be mapped to a dense vector with a fixed dimension d_model (512 in this example) through a trainable linear projection layer. This projection operation is part of the encoder and is learned synchronously during model training. For the sake of simplicity, the vector with dimension d_model after this projection transformation will still be called the weighted keyword vector and input into subsequent positional encoding steps.

[0027] The position augmentation vector sequence is input into a deep semantic encoder. This encoder is trained in a supervised manner during the model building phase using historical application material texts and their manually labeled category tags. The training data comes from a de-identified and labeled corpus of historical application materials. The encoder consists of six stacked layers with identical structures. The number of layers (6) was determined by a grid search on the validation set during the model building phase, with candidate values ​​including 4, 6, and 8. The layer number that achieved the highest classification accuracy on the validation set was ultimately selected. Each encoding layer contains a multi-head self-attention sublayer and a feedforward fully connected network sublayer. The number of heads (h) in the multi-head self-attention sublayer is 8, determined by a grid search on the validation set for candidate values ​​(e.g., 8, 12, 16) based on the best classification performance. The internal hidden layer dimension (d_ff) of the feedforward sublayer is 2048. This dimension was determined by a grid search on the validation set during the model building phase. The specific process involves training the model using historical application material text and its labeled categories. Different candidate values ​​for d_ff are tested on the validation set. These candidate values ​​are generated based on multiples of d_model, including d_model multiplied by 2, d_model multiplied by 4, and d_model multiplied by 8. The system evaluates the classification accuracy of the model corresponding to each candidate value on the validation set, and finally selects the candidate value d_model multiplied by 4 that achieves the highest classification accuracy as the fixed value of d_ff. The training process is completed by minimizing the cross-entropy loss function using the Adam optimizer. In the current processing, the position augmentation vector sequence passes through six encoding layers sequentially, undergoing contextual information aggregation and nonlinear transformations, ultimately outputting the hidden layer state tensor.

[0028] The system applies a multi-head self-attention mechanism to the hidden layer state tensor again. This mechanism uniformly divides the hidden layer state tensor into eight heads along the feature dimension. For each head, the corresponding trainable weight matrix is ​​used to linearly project the input to generate a query matrix, a key matrix, and a value matrix. Attention weights are obtained by calculating the dot product of the query matrix and the transpose of the key matrix, dividing by a scaling factor of 8, and then applying a normalized exponential function. The value matrix is ​​then weighted and summed using these weights to obtain the output of that head. The outputs of the eight heads are concatenated along the feature dimension to form a merged tensor. This merged tensor is transformed by a learnable linear projection weight matrix W_O, with dimensions 4096 x 512. The parameters of this matrix are randomly initialized at the beginning of model training and updated and optimized through gradient backpropagation during training. The output after the projection transformation is the context feature vector. This step aims to capture and fuse diverse lexical association patterns from different semantic subspaces.

[0029] The system calculates the variance of the context feature vector in each feature dimension to obtain the context feature variance. The preset semantic discrete threshold is determined based on the distribution of context feature variances generated by all training samples during the encoder training phase. Specifically, after model training is complete, all training samples are forward-propagated once to collect all generated context feature variance values. This variance value sequence is then sorted in ascending order, and the value at the 90th percentile is taken as the semantic discrete threshold. If the context feature variance calculated from the current application material is not greater than this threshold, the context feature vector is directly determined as the multidimensional semantic representation vector.

[0030] If the variance of the context features exceeds this threshold, the system performs vector reconstruction based on metric mapping. The algorithm first needs to determine the number of neighbors, *k*, and the target low-dimensional dimension, *d_re*. The number of neighbors, *k*, is determined through a grid search on the validation set during the model building phase, with candidate values ​​[5, 10, 15, 20] selected to maximize the average silhouette coefficient in subsequent clustering steps on the validation set. The target low-dimensional dimension, *d_re*, is also determined through a grid search and cross-validation, selecting the dimension from candidate values ​​[32, 64, 128, 256] that maintains optimal class separation in the dimensionality-reduced data on the validation set. After determining the parameters, the algorithm calculates the Euclidean distance between samples based on the context feature vectors, selects the k nearest points for each sample as neighbors to construct an adjacency graph, and calculates the shortest path distance between all pairs of points in the graph as an approximation of the geodesic distance. The algorithm then performs eigenvalue decomposition on the centered inner product matrix obtained from the geodesic distance matrix transformation, retaining the eigenvectors corresponding to the first *d_re* largest eigenvalues, and projects the original data onto this low-dimensional space to obtain a low-dimensional manifold vector.

[0031] Finally, a linear interpolation algorithm is used to reconstruct the low-dimensional manifold vector of dimension d_re to the original dimension of 512. The interpolation algorithm uniformly inserts new points within the index interval of the low-dimensional vector. The value of each new point is linearly calculated from the values ​​of its two adjacent original data points, until the vector length reaches 512. The reconstructed vector is the final multidimensional semantic representation vector. This adaptive reconstruction process handles complex texts with highly discrete semantics, transforming high-dimensional sparse features into a low-dimensional, compact semantic manifold representation, thus improving the discriminability of the semantic representation.

[0032] In step S14, based on the multidimensional semantic representation vector and the preset semantic category standard vector, the semantic similarity between vector directions is calculated one by one to obtain a semantic similarity sequence, including: When the multidimensional semantic representation vector is inconsistent with the vector dimension of the preset category standard vector library, the principal component analysis algorithm is used to perform dimensionality reduction and spatial projection to obtain the vector to be matched. Based on the vector to be matched, extract the centroid vector of each category from the category standard vector library, calculate the semantic similarity of each category using the cosine similarity algorithm, and multiply by a preset category weight factor to obtain a semantic similarity sequence.

[0033] Specifically, based on the multidimensional semantic representation vector output in step S13, a vector dimension consistency check is first performed. A pre-defined category standard vector library is invoked, which stores the standard vector corresponding to each predefined material category. This category standard vector library is constructed during system initialization, and its data originates from historically accurately classified application material samples. The construction process involves collecting all historical samples belonging to each category, generating a multidimensional semantic representation vector for each sample using steps S11 to S13 (same as the current process), and then calculating the arithmetic mean of these vectors in each dimension. The resulting mean vector is the centroid vector of the category and is stored in the vector library as the semantic category standard vector for that category.

[0034] The system compares the dimension of the current multidimensional semantic representation vector with the dimension of vectors in the category standard vector library. The dimension of the current multidimensional semantic representation vector is fixed at d_model (i.e., 512). The dimension of the category standard vector library is consistent with the d_model used in historical training; therefore, the dimension should be the same when the system is running normally and the version is consistent. If the dimension is inconsistent due to model version iteration, i.e., the dimension of the current vector is not equal to the dimension of the vector in the library, the system starts the principal component analysis algorithm for dimensionality reduction and alignment. This algorithm uses all centroid vectors in the category standard vector library as input data, calculates the covariance matrix of these data, and performs eigenvalue decomposition on the matrix. The eigenvectors corresponding to the largest d_target eigenvalues ​​are selected to form a projection matrix, and d_target is set as the actual dimension of the current multidimensional semantic representation vector. Then, the current multidimensional semantic representation vector is linearly projected using this projection matrix to reduce it to the d_target dimension, obtaining the vector to be matched. If the dimensions are consistent, the multidimensional semantic representation vector is directly used as the vector to be matched.

[0035] The system then calculates the semantic similarity between the vector to be matched and the centroid vector of each category. This process iterates through the centroid vectors of each category in the standard category vector library. For each category, the cosine similarity between the vector to be matched and the centroid vector of that category is calculated. The cosine similarity is calculated by first calculating the dot product of the two vectors, then calculating the Euclidean norm of each vector (i.e., the square root of the sum of the squares of the values ​​in each dimension), and finally dividing the dot product by the product of the two norms. The result is the similarity value, ranging from -1 to +1.

[0036] After obtaining the original cosine similarity value for each category, the system needs to multiply it by a preset category weighting factor. The category weighting factor is determined during the construction of the vector library. Its calculation is based on the frequency of each category's occurrence in the historical sample library. Specifically, the steps are as follows: count the number of samples for each category in the historical sample library, calculate the proportion of each category's samples to the total number of samples, and then normalize these proportions so that the sum of the weighting factors for all categories is 1. This weighting factor is used to adjust the influence of prior category probabilities on similarity, preventing high-frequency categories from having an excessive advantage in similarity comparisons. The original cosine similarity value for each category is multiplied by its corresponding category weighting factor to obtain the adjusted semantic similarity value. The adjusted semantic similarity values ​​for all categories are arranged in order of their category index, thus forming a semantic similarity sequence.

[0037] This step generates a numerical sequence that comprehensively reflects the relevance of the material to be classified to each category by comparing the semantic representation of the material with the standard semantic centers of all preset categories and quantifying their directional similarity. This semantic similarity sequence is the direct basis for subsequent threshold judgments and classification decisions, and its accuracy is directly related to the reliability of the entire classification process.

[0038] In step S15, based on the semantic similarity sequence and a preset semantic high similarity threshold, a material type matching analysis is performed to determine the preliminary material classification result, including: Based on the semantic similarity sequence, the highest semantic similarity value is found by traversing the sequence. When the highest semantic similarity value is greater than the preset semantic high similarity threshold, the position number corresponding to the highest semantic similarity value in the semantic similarity sequence is determined as the target index identifier; When the highest semantic similarity value is not greater than a preset high semantic similarity threshold, the second highest semantic similarity value is extracted from the semantic similarity sequence, and the absolute difference between the second highest semantic similarity value and the highest semantic similarity value is calculated to obtain the semantic similarity deviation. When the semantic similarity deviation is less than or equal to a preset semantic deviation threshold, the position number corresponding to the second highest semantic similarity value in the semantic similarity sequence is determined as the target index identifier. When the semantic similarity deviation is greater than the preset semantic deviation threshold, it is determined that the application material text cannot be directly classified through this round of matching, and a preliminary classification result of the material containing the material unmatched identifier is obtained; Based on the target index identifier, the corresponding category label is retrieved and obtained by querying the index-label mapping relationship pre-stored in the category standard vector library, the material category of the application material text is determined, and a preliminary material classification result containing the material matching identifier and the material category is obtained.

[0039] Specifically, based on the semantic similarity sequence output in step S14, material type matching analysis begins. First, all values ​​in the semantic similarity sequence are iterated through, and the maximum value is found through sequential comparison; this value is recorded as the highest semantic similarity value. The preset semantic high similarity threshold is determined based on historical classification data statistics. The specific setting steps are as follows: before system deployment, test samples of correctly classified application materials are used, and the highest semantic similarity value generated by each sample during the classification process is recorded. After collecting all these highest values, they are sorted in ascending order of numerical value, and the value located at the 85th percentile is taken as the semantic high similarity threshold. This threshold means that historically, 85% of clearly classified samples have a similarity higher than this value.

[0040] If the highest semantic similarity value of the current application material is greater than the high semantic similarity threshold, the system determines that the material has a sufficiently high confidence level in matching a certain category. The system then determines the position number of this highest value in the semantic similarity sequence, which directly corresponds to the index of a certain category in the category standard vector library. This position number is then used as the target index identifier.

[0041] If the highest semantic similarity value is not greater than the high semantic similarity threshold, the system initiates an ambiguity resolution process. This process first extracts the second-highest value from the semantic similarity sequence, i.e., the second-highest semantic similarity value. The absolute difference between the second-highest and highest semantic similarity values ​​is calculated to obtain the semantic similarity deviation. The system compares this deviation value with a preset semantic deviation threshold. The semantic deviation threshold is also determined based on historical data; it collects all semantic similarity deviation values ​​from historical classification samples where the highest value is not greater than the high threshold, and takes the 90th percentile of these values ​​in ascending order as the threshold.

[0042] When the semantic similarity deviation is less than or equal to the semantic deviation threshold, it indicates that the categories corresponding to the highest and second-highest values ​​are in fierce competition, and the material's semantics may lie between the two. In this case, the system chooses to determine the position number corresponding to the second-highest semantic similarity value in the sequence as the target index identifier. This design tends to select the relatively more suitable suboptimal category in ambiguous situations, rather than forcibly using the highest value category with insufficient confidence.

[0043] When the semantic similarity deviation exceeds the semantic deviation threshold, the system determines that the semantic representation of the current application material text differs significantly from the standard vector of any preset category, or lies between multiple categories but lacks a clear tendency, making reliable direct classification impossible through this round of matching. In this case, the system generates a preliminary classification result containing a specific material mismatch identifier, which is used to trigger different processing branches in subsequent steps.

[0044] For the first two scenarios where the target index identifier is successfully determined, the system needs to convert it into readable category information. The system queries the pre-stored index-label mapping table in the category standard vector library. This mapping table is generated during the construction of the category standard vector library and stores the correspondence between the storage index number of each category centroid vector and its corresponding text category label (such as "work injury determination application," "enterprise registration change," etc.). The system uses the target index identifier to retrieve this table, obtains the corresponding category label, and thus determines the material category of the application material text. Finally, the system generates a preliminary material classification result containing the material matching identifier and the determined material category.

[0045] This step, by setting a high confidence threshold and a bias analysis mechanism, achieves preliminary quality control of the classification results. Materials that can be clearly classified are quickly output, while materials with insufficient matching or ambiguity are identified, providing clear input for subsequent enhancement processing and fine-grained classification. This avoids the direct output of low-quality classification results and ensures the accuracy of the final classification results from a process perspective.

[0046] In step S16, when the preliminary material classification result shows that the type match is successful, the material type is output as the final classification result.

[0047] Specifically, the preliminary material classification result is received as the step input. This result is a structured data object containing two key fields: material matching identifier and material category. The material matching identifier is a boolean value, which indicates that a type match was successfully determined in step S15 when it is true. The material category is a string corresponding to the category label retrieved from the category standard vector library.

[0048] When the material matching flag in the preliminary material classification result is true, the success output process is triggered. The material category field in the preliminary material classification result is extracted. This content represents the final category assignment determined by the system for the current application material text after keyword extraction, semantic representation, similarity calculation, and threshold matching.

[0049] The material category string is then bound to the unique application identifier of the currently processed application material text, generating a final classification result record. This record is written to the system's pre-set final classification result database. This database is created during system initialization to persistently store the classification conclusions of all processed materials, and its table structure includes fields for application identifier, material category, and classification time. The write operation ensures data consistency, meaning that the same application identifier corresponds to only one latest final classification record in the database.

[0050] In step S17, when the preliminary classification result of the material shows a type matching failure, the distinguishing weight of the keywords is re-evaluated based on the application material text, and the multidimensional semantic representation vector is supplemented and strengthened based on the re-evaluated distinguishing weight to form an enhanced semantic representation vector, including: By capturing adjacent sentences of the keywords from the application material text through a preset sliding window, a semantically related text set is obtained; Based on the semantically related text set, the global salient weight of each word is calculated using the term frequency-inverse document frequency algorithm, generating a weight score sequence corresponding to the semantically related text. Based on the semantically related text set and the corresponding weight score sequence, using the global vocabulary as the index, each word appearing in the semantically related text is assigned a weight score to the corresponding dimension position in the vector, while the dimension of words that do not appear is set to zero, thus generating a word dimension vector. The word dimension vectors are then stacked vertically according to the order of the semantically related texts to construct a word vector feature matrix. When the word vector feature matrix exceeds a preset dimension threshold, dimensionality reduction is performed using principal component analysis to obtain a dimensionality-reduced word vector feature matrix. When the word vector feature matrix does not exceed the preset dimension threshold, the original word vector feature matrix is ​​obtained directly; Based on the word vector feature matrix and the multidimensional semantic representation vector, a weighted fusion process is performed to obtain the enhanced semantic representation vector.

[0051] Specifically, when the material matching identifier in the preliminary material classification result is false, the type matching is deemed to have failed, and the enhanced semantic representation process is triggered. This process first re-evaluates the context weight of the keywords. Starting from the original application material text, for each keyword extracted in step S12, a sliding window of fixed size K characters is used for context capture. The center of the window is aligned with the starting position of the keyword in the text, and K / 2 characters are truncated forward and backward, stopping when a text boundary is encountered. The truncated character fragments containing the keyword and its preceding and following context are considered as a semantically related text. All semantically related texts corresponding to the keywords together constitute a semantically related text set. The sliding window size K is determined based on the length distribution of sentences in historical application materials, specifically taking the median length of all sentences in the historical corpus. Specifically, the sentence length refers to the number of lexical units contained in the sentence after Chinese word segmentation, and the value is the median length of all sentences in the historical corpus. The aim is to obtain a window size that can cover the core part of a typical sentence. In actual implementation, K can also be set to a fixed value (such as 200 characters), or dynamically determined according to the principle of several complete lexical units adjacent to the keyword (such as 5-10 words before and after). The core purpose is to capture the local contextual semantic fragments of the keyword.

[0052] Subsequently, a new weight for each word is calculated based on the semantically related text set. This calculation uses a term frequency-inverse document frequency (IF-IVF) algorithm, with IVF statistics based on a global corpus of historical application materials. For each text segment in the semantically related text set, the term frequency weight of each word is calculated, which is the number of times the word appears in the segment divided by the total number of words in the segment. Simultaneously, the IVF weight of the word is obtained from the pre-stored global corpus document frequency statistics. Multiplying the term frequency weight by the IVF weight yields the global salient weight of the word in that text segment. A weight score sequence consisting of the words it contains and their corresponding global salient weights is generated for each semantically related text.

[0053] Next, the word vector feature matrix is ​​constructed. This process relies on a pre-built global vocabulary, generated during system initialization. This vocabulary contains all unique words that have appeared in the historical global corpus, and each word is assigned a unique dimension index. For each semantically related text and its weight score sequence, the system creates a zero vector with a length equal to the total number of words in the global vocabulary. The weight score sequence of the text is traversed. For each word in the sequence, its corresponding position in the zero vector is found based on its index in the global vocabulary, and the global salient weight value of that word is assigned to that position. The vector dimension corresponding to words that do not appear in the text remains zero. This vector is the word dimension vector of the semantically related text. Following the original order of the semantically related text set, all generated word dimension vectors are stacked vertically to form the word vector feature matrix, where each row represents a text segment, and each column corresponds to a word in the global vocabulary.

[0054] The system determines whether the dimensionality of the word vector feature matrix exceeds a preset dimensionality threshold. This threshold is set as the total number of words in the global vocabulary multiplied by a preset scaling factor R. The scaling factor R is determined as follows: During the system construction phase, a subset of samples from historical data is used to test the matrix at multiple candidate R values ​​(e.g., 0.05, 0.1, 0.15, 0.2). The average accuracy of the dimensionality-reduced matrix corresponding to different R values ​​on subsequent classification validation tasks is evaluated, and the candidate R value with the highest average accuracy is selected as the final fixed value. If the number of columns in the matrix exceeds the threshold, principal component analysis (PCA) is used for dimensionality reduction. The target dimensionality reduction dimension d_pca is set to the size of the dimensionality threshold, i.e., the integer part of the product of the total number of words in the global vocabulary and R. The dimensionality-reduced word vector feature matrix is ​​obtained after dimensionality reduction. If the threshold is not exceeded, the original word vector feature matrix is ​​used directly.

[0055] Finally, the word vector feature matrix and the multidimensional semantic representation vector generated in step S13 are weighted and fused. The multidimensional semantic representation vector is first copied and expanded into a matrix with the same number of rows as the word vector feature matrix, and each row contains the same content. Weighted fusion uses element-wise weighted summation, assigning fusion weight factors α and β to the two matrices. The weight factors α and β are determined through optimization using a historical augmentation sample set. Specifically, a grid search is performed on the historical validation set, with α and β in the range of 0 to 1 with a fixed step size. Cross-validation is then used to select the weight combination that maximizes the final classification accuracy on the validation set as fixed parameters. The word vector feature matrix is ​​multiplied by the weight factor α, and the expanded multidimensional semantic representation vector matrix is ​​multiplied by the weight factor β. The two resulting matrices are then summed. Finally, mean pooling is performed on the summed matrix in the row direction, i.e., the arithmetic mean of each column is calculated, resulting in a single vector, which is the augmented semantic representation vector. This step effectively supplements and corrects the initial semantic representation by incorporating richer local contextual information, providing a more discriminative feature representation for materials that are difficult to classify directly.

[0056] In step S18, based on the enhanced semantic representation vector, a new semantic similarity sequence is obtained by calculating the semantic similarity between the enhanced semantic representation vector and the semantic category standard vector. This sequence is then combined with a preset low semantic similarity threshold to filter out abnormal application materials, including: Based on the enhanced semantic representation vector and the category centroid vector, new semantic similarity is calculated for each category using the cosine similarity algorithm, and then multiplied by the category weight factor to obtain a new semantic similarity sequence. Obtain the highest new semantic similarity value of the new semantic similarity sequence. When the highest new semantic similarity value is greater than a preset low semantic similarity threshold, determine the position number corresponding to the highest new semantic similarity value in the new semantic similarity sequence as the target index identifier. Based on the target index identifier, the corresponding category label is retrieved and obtained by querying the index-tag mapping relationship pre-stored in the category standard vector library, thereby determining the material category of the application material text; When the highest value of the new semantic similarity is less than or equal to the preset low semantic similarity threshold, it is determined that the application material text cannot be reliably classified, and an abnormal classification mark is added to the application material text. Based on the application material text and all pre-stored source text data carrying the anomaly classification marker, the abnormal application materials are identified by summarizing through association identifier retrieval.

[0057] Specifically, the enhanced semantic representation vector is obtained by contextually supplementing and strengthening the original multidimensional semantic representation vector in step S17. This is achieved by accessing a pre-defined category standard vector library, which is built during system initialization and stores the centroid vector of each predefined category (i.e., the semantic category standard vector) and its associated category weight factor. The category weight factor is calculated based on the historical sample frequency of each category, specifically the proportion of samples from each category to the total number of historical samples, and is normalized to a sum of one.

[0058] Based on the enhanced semantic representation vector and the centroid vector of each category in the standard category vector library, semantic similarity is calculated class-by-class using the cosine similarity algorithm. The calculation process is as follows: for each category, first, the dot product of the enhanced semantic representation vector and the centroid vector of that category is calculated; then, the Euclidean norm of each vector is calculated; finally, the dot product result is divided by the product of the two norms to obtain the original cosine similarity value. This original similarity value is multiplied by the category weight factor corresponding to that category to obtain the adjusted new semantic similarity value. After traversing all categories, these new semantic similarity values ​​are arranged in order of their category index, thus forming a new semantic similarity sequence.

[0059] The new semantic similarity sequence is then iterated through to find the maximum value, which is recorded as the highest new semantic similarity value. The preset semantic low similarity threshold is determined based on historical classification data. Specifically, the steps are as follows: collect all samples that were correctly classified after enhancement processing in the past, and record their highest new semantic similarity values. These highest values ​​are then sorted in ascending order, and the value at the 20th percentile is taken as the semantic low similarity threshold. This threshold is lower than the high threshold in step S15, aiming to provide a more lenient matching threshold for enhanced materials.

[0060] In another implementation, to accurately filter out anomalous materials with extremely low semantic matching, the semantic low similarity threshold should be set based on samples that still fail to be classified after augmentation, or by comprehensively considering the tail of the similarity distribution of all samples. This can be achieved by collecting the highest new semantic similarity values ​​generated after augmentation of all historical samples (including those that were successfully and unsuccessfully classified), arranging all these values ​​in ascending order, and taking the value at or below the 10th percentile as the semantic low similarity threshold. This method ensures that only materials with similarity at the very end of the overall distribution are judged as anomalous, thereby improving the strictness and accuracy of anomalous screening.

[0061] If the highest new semantic similarity value of the current material is greater than the semantic low similarity threshold, the system determines that the enhanced material can be classified. The system determines the position number of this highest value in the new semantic similarity sequence and uses this number as the target index identifier. Next, the system queries the pre-stored index-label mapping table in the category standard vector library. This mapping table is generated when the vector library is built and stores the correspondence between the storage index of each category centroid vector and the text category label. Using the target index identifier, this table is retrieved to obtain the corresponding category label, thereby determining the material category of the application material text.

[0062] If the highest new semantic similarity value is less than or equal to the low semantic similarity threshold, the system determines that even after semantic enhancement, the semantic match between the application material text and any known category is still too low to be reliably classified. In this case, the system adds an anomalous classification tag to the metadata of the application material text. This tag is a specific attribute identifier used to quickly identify such materials in subsequent processes.

[0063] Based on the application materials currently marked with an anomaly classification, and combined with all historical source text data carrying this mark pre-stored in the system, a summary is generated through association identifier retrieval. The association identifier is a unique identifier for each application material. The resulting set is the set of anomaly application materials for the current batch. This set will be passed as output to step S19 for final classification output.

[0064] The significance of this step lies in setting up a second classification filter for the semantically enhanced material. By using a more lenient, lower threshold, it attempts to capture material that failed in the initial classification but may be clearly categorized after enhancement. Simultaneously, it rigorously filters out "truly anomalous" material that exhibits significant semantic differences from all standard categories, accurately identifying and summarizing them. This provides clear processing targets for the final fine-grained classification based on clustering and consistency assessment, ensuring that the system's final output has both high coverage and high reliability.

[0065] In step S19, based on the abnormal application materials, semantic proximity analysis is performed, and the materials are grouped according to the semantic proximity analysis results through semantic consistency evaluation to obtain the final classification results, including: Based on the abnormal application materials, the weight of each keyword is calculated using the term frequency-inverse document frequency algorithm, and a feature weight distribution matrix is ​​constructed with the application materials as rows and the keyword weights as columns; Based on the feature weight distribution matrix, the DBSCAN density clustering algorithm is used to aggregate the abnormal application materials whose mutual distance is less than the preset neighborhood radius and whose number of materials around the core point reaches the preset minimum sample number requirement, using Euclidean distance as the measure of dissimilarity between vectors, to form semantic proximity groups. Calculate the arithmetic mean of the enhanced semantic representation vector corresponding to the abnormal application materials contained in the semantic proximity group in each dimension to generate the overall semantic vector; Based on the overall semantic vector, the overall similarity between the overall semantic vector and the enhanced semantic representation vector is calculated using the cosine similarity algorithm, and the arithmetic mean of the overall similarity is obtained as the average similarity evaluation index of the semantic proximity group. Based on the average similarity evaluation index, material category labels are matched in the category standard vector library, and the material category labels are assigned to all the abnormal application materials within the semantic proximity group to determine the final classification result.

[0066] Specifically, the set of anomalous application materials includes all application material texts that were tagged with anomalous classification in step S18 and their corresponding enhanced semantic representation vectors. First, a feature weight distribution matrix is ​​constructed for each anomalous application material in the set. This process is based on the keyword list extracted for each material in step S12. For each keyword in the list, its weight is recalculated using a term frequency-inverse document frequency (IF-IVF) algorithm, where IVF statistics are based on a historical global corpus. The calculated weight values ​​reflect the importance of the keyword in a specific anomalous material. All anomalous application materials are arranged as rows, and all appearing keywords are arranged as columns. Each cell is filled with the weight value of the corresponding keyword in the material; if a keyword does not appear in a material, it is filled with zero, thus constructing the feature weight distribution matrix.

[0067] Subsequently, the DBSCAN density clustering algorithm is used to perform cluster analysis on the feature weight distribution matrix. This algorithm requires two preset parameters: neighborhood radius eps and minimum sample size minPts. The neighborhood radius eps is determined as follows: during the system construction phase, the Euclidean distances between all historical anomalous samples in the feature weight space are calculated, a k-distance curve is plotted (k value is minPts), and the distance value corresponding to the inflection point of the curve is selected as eps. The minimum sample size minPts is set based on the principle that an effective semantic group should contain at least a certain number of materials. This value is set as the total number of historical anomalous material sets multiplied by a fixed proportion (e.g., 0.02) and then rounded up. The algorithm runs as follows: first, the Euclidean distance between each row vector (i.e., each material) in the feature weight distribution matrix is ​​calculated, and all core points that satisfy the condition of having at least minPts sample points in their neighborhood with a radius of eps are found. Then, each core point is connected to all points in its neighborhood (including other core points and boundary points) to form density-connected clusters. All points not included in the neighborhood of any core point are marked as noise points. Finally, the algorithm outputs several semantic neighbor groups that aggregate core points and boundary points together, while noise points are temporarily kept in a separate set.

[0068] For each semantic proximity group, semantic consistency is assessed. First, the enhanced semantic representation vectors corresponding to all anomalous application materials within the group are extracted, and the arithmetic mean of these vectors on each feature dimension is calculated to generate a global semantic vector representing the overall semantic tendency of the group. Next, the cosine similarity between the enhanced semantic representation vector of each material within the group and the global semantic vector is calculated. The arithmetic mean of all calculated cosine similarity values ​​is then taken to obtain the average similarity assessment index for the group. This index is used to quantify the degree of semantic consistency within the group.

[0069] Finally, a final category label is assigned to each semantic proximity group. This process is accomplished by calculating the cosine similarity between the overall semantic vector corresponding to the group and the centroid vector of each category in the category standard vector library. The category centroid vector with the highest similarity is selected, and its corresponding category label is assigned to all anomalous application materials within that semantic proximity group. If the highest similarity is lower than a preset matching threshold (set as the 10th percentile of the similarity distribution of historical normal classification samples), the group and its materials will remain in the "undetermined" category and be output for manual review. The classification results of all anomalous application materials that have been successfully assigned category labels are combined with the results output in step S16 to constitute the final classification result of the system.

[0070] This step automatically identifies and aggregates anomalous materials with ambiguous semantic features that cannot be directly matched with standard categories but exhibit high semantic similarity. Then, through intra-group consistency assessment and secondary matching of the group as a whole with standard categories, these materials are assigned appropriate category labels. This mechanism effectively handles marginal cases and novel expression patterns that traditional methods struggle to address, improving the system's classification coverage and accuracy for complex and unconventional application materials.

[0071] Reference Figure 2 The second embodiment of the present invention provides an intelligent classification system for application materials based on data analysis, comprising: The data acquisition module is used to obtain the text of application materials; The keyword analysis module is used to extract keywords, calculate and weight them according to the application material text, and obtain a weighted keyword vector. The semantic feature analysis module is used to map the weighted keyword vector to a high-dimensional semantic space by capturing the contextual associations between words, based on the weighted keyword vector, to obtain a multi-dimensional semantic representation vector. The semantic similarity analysis module is used to calculate the semantic similarity between vector directions one by one based on the multidimensional semantic representation vector and the preset semantic category standard vector, so as to obtain a semantic similarity sequence. The preliminary classification module is used to perform material type matching analysis based on the semantic similarity sequence and a preset semantic high similarity threshold to determine the preliminary material classification result; The preliminary output module is used to output the material type as the final classification result when the preliminary classification result of the material shows that the type match is successful. The enhanced semantic module is used to re-evaluate the distinguishing weight of keywords based on the application material text when the preliminary classification result of the material shows that the type matching fails, and to supplement and strengthen the multi-dimensional semantic representation vector based on the re-evaluated distinguishing weight to form an enhanced semantic representation vector. The anomaly screening module is used to obtain a new semantic similarity sequence by calculating the semantic similarity between the enhanced semantic representation vector and the semantic category standard vector based on the enhanced semantic representation vector, and to screen out abnormal application materials in combination with a preset semantic low similarity threshold. The final classification output module is used to perform semantic proximity analysis on the abnormal application materials, and group them according to the semantic proximity analysis results through semantic consistency evaluation to obtain the final classification results.

[0072] It should be noted that the intelligent classification system for application materials based on data analysis provided in this embodiment of the invention is used to execute all the process steps of the intelligent classification method for application materials based on data analysis in the above embodiments. The working principles and beneficial effects of the two correspond one-to-one, so they will not be described again.

[0073] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a data analysis-based intelligent classification program for application materials. When the processor executes the computer program, it implements the steps in the various data analysis-based intelligent classification method embodiments described above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the intelligent classification module for application materials based on data analysis.

[0074] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0075] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0076] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0077] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0078] If the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0079] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0080] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A data analysis-based intelligent classification method for application materials, characterized in that, include: Obtain the application materials text; Based on the application material text, keyword extraction, differentiation weight calculation, and weighting are performed to obtain a weighted keyword vector. This includes: segmenting the text and extracting vocabulary units using Chinese word segmentation technology to obtain word segmentation results; calculating the product of the word frequency weight and inverse document frequency weight of each vocabulary unit using a word frequency-inverse document frequency (IF-IVF) algorithm based on the word segmentation results and pre-stored corpus document frequency statistics to generate a keyword weight sequence; calculating the statistical variance of the frequency of each vocabulary unit in the application material text based on the keyword weight sequence to obtain the differentiation weight of each vocabulary unit; weighting and fusing the word frequency weight, the IVF weight, and the differentiation weight to obtain a key feature vector; selecting key feature vectors greater than a preset key feature threshold to form an optimized feature set; and performing normalization and vector dimension alignment based on the optimized feature set to obtain a weighted keyword vector. Based on the weighted keyword vector, the weighted keyword vector is mapped to a high-dimensional semantic space by capturing the contextual associations between words to obtain a multi-dimensional semantic representation vector. Based on the multidimensional semantic representation vector and the preset semantic category standard vector, the semantic similarity between vector directions is calculated one by one to obtain a semantic similarity sequence; Based on the semantic similarity sequence and the preset semantic high similarity threshold, material type matching analysis is performed to determine the preliminary material classification result; When the preliminary classification result of the material shows that the type match is successful, the material type is output as the final classification result. When the initial classification result of the material shows that the type matching fails, the distinguishing weight of the keywords is re-evaluated according to the application material text, and the multidimensional semantic representation vector is supplemented and strengthened according to the re-evaluated distinguishing weight to form an enhanced semantic representation vector. Based on the enhanced semantic representation vector, a new semantic similarity sequence is obtained by calculating the semantic similarity between the enhanced semantic representation vector and the semantic category standard vector, and abnormal application materials are screened out in combination with a preset semantic low similarity threshold. Based on the abnormal application materials, semantic proximity analysis is performed, and the groups are grouped according to the semantic proximity analysis results through semantic consistency evaluation to obtain the final classification results.

2. The intelligent classification method for application materials based on data analysis according to claim 1, characterized in that, The step of mapping the weighted keyword vector to a high-dimensional semantic space by capturing the contextual relationships between words, based on the weighted keyword vector, to obtain a multi-dimensional semantic representation vector includes: Based on the weighted keyword vector, calculate the sine and cosine position encoding vector for each word, and add the sine and cosine position encoding vector to the weighted keyword vector to obtain a position enhancement vector sequence; The position enhancement vector sequence is input into a pre-built deep semantic encoder based on a self-attention mechanism. The encoder performs nonlinear transformation and context encoding through a multi-layer self-attention network and a fully connected network to output the hidden layer state tensor. Based on the hidden layer state tensor, the association strength of lexical features in different semantic subspaces is calculated through a multi-head self-attention mechanism, and the position enhancement vector sequence is weighted and fused to obtain the context feature vector. Semantic discrete judgment is performed based on the context feature vector, and vector reconstruction is performed based on the semantic discrete judgment result to obtain a multidimensional semantic representation vector.

3. The intelligent classification method for application materials based on data analysis according to claim 2, characterized in that, The step of performing semantic discrete judgment based on the context feature vector and reconstructing the vector based on the semantic discrete judgment result to obtain a multi-dimensional semantic representation vector includes: Calculate the numerical variance of the context feature vector in each dimension to obtain the context feature variance; When the variance of the context features is not greater than a preset semantic discrete threshold, the context feature vector is directly determined as a multidimensional semantic representation vector. When the variance of the context features is greater than the preset semantic discrete threshold, nonlinear dimensionality reduction and cluster separation are performed based on the context feature vector using an equal metric mapping algorithm, and the low-dimensional manifold vector obtained after dimensionality reduction is reconstructed to the original dimension to obtain a multidimensional semantic representation vector.

4. The intelligent classification method for application materials based on data analysis according to claim 1, characterized in that, The step of calculating the semantic similarity between vector directions one by one based on the multidimensional semantic representation vector and the preset semantic category standard vector to obtain a semantic similarity sequence includes: When the multidimensional semantic representation vector is inconsistent with the vector dimension of the preset category standard vector library, the principal component analysis algorithm is used to perform dimensionality reduction and spatial projection to obtain the vector to be matched. Based on the vector to be matched, extract the centroid vector of each category from the category standard vector library, calculate the semantic similarity of each category using the cosine similarity algorithm, and multiply by a preset category weight factor to obtain a semantic similarity sequence.

5. The intelligent classification method for application materials based on data analysis according to claim 4, characterized in that, The step of performing material type matching analysis based on the semantic similarity sequence and a preset semantic high similarity threshold to determine the preliminary material classification result includes: Based on the semantic similarity sequence, the highest semantic similarity value is found by traversing the sequence. When the highest semantic similarity value is greater than the preset semantic high similarity threshold, the position number corresponding to the highest semantic similarity value in the semantic similarity sequence is determined as the target index identifier; When the highest semantic similarity value is not greater than a preset high semantic similarity threshold, the second highest semantic similarity value is extracted from the semantic similarity sequence, and the absolute difference between the second highest semantic similarity value and the highest semantic similarity value is calculated to obtain the semantic similarity deviation. When the semantic similarity deviation is less than or equal to a preset semantic deviation threshold, the position number corresponding to the second highest semantic similarity value in the semantic similarity sequence is determined as the target index identifier. When the semantic similarity deviation is greater than the preset semantic deviation threshold, it is determined that the application material text cannot be directly classified through this round of matching, and a preliminary classification result of the material containing the material unmatched identifier is obtained; Based on the target index identifier, the corresponding category label is retrieved and obtained by querying the index-label mapping relationship pre-stored in the category standard vector library, the material category of the application material text is determined, and a preliminary material classification result containing the material matching identifier and the material category is obtained.

6. The intelligent classification method for application materials based on data analysis according to claim 1, characterized in that, When the initial classification result of the material shows a type matching failure, the distinguishing weight of the keywords is reassessed based on the application material text, and the multidimensional semantic representation vector is supplemented and strengthened based on the reassessed distinguishing weight to form an enhanced semantic representation vector, including: By capturing adjacent sentences of the keywords from the application material text through a preset sliding window, a semantically related text set is obtained; Based on the semantically related text set, the global salient weight of each word is calculated using the term frequency-inverse document frequency algorithm, generating a weight score sequence corresponding to the semantically related text. Based on the semantically related text set and the corresponding weight score sequence, using the global vocabulary as the index, each word appearing in the semantically related text is assigned a weight score to the corresponding dimension position in the vector, while the dimension of words that do not appear is set to zero, thus generating a word dimension vector. The word dimension vectors are then stacked vertically according to the order of the semantically related texts to construct a word vector feature matrix. When the word vector feature matrix exceeds a preset dimension threshold, dimensionality reduction is performed using principal component analysis to obtain a dimensionality-reduced word vector feature matrix. When the word vector feature matrix does not exceed the preset dimension threshold, the original word vector feature matrix is ​​obtained directly; Based on the word vector feature matrix and the multidimensional semantic representation vector, a weighted fusion process is performed to obtain the enhanced semantic representation vector.

7. The intelligent classification method for application materials based on data analysis according to claim 4, characterized in that, The step involves calculating the semantic similarity between the enhanced semantic representation vector and the semantic category standard vector to obtain a new semantic similarity sequence, and then filtering out abnormal application materials based on a preset low semantic similarity threshold, including: Based on the enhanced semantic representation vector and the category centroid vector, new semantic similarity is calculated for each category using the cosine similarity algorithm, and then multiplied by the category weight factor to obtain a new semantic similarity sequence. Obtain the highest new semantic similarity value of the new semantic similarity sequence. When the highest new semantic similarity value is greater than a preset low semantic similarity threshold, determine the position number corresponding to the highest new semantic similarity value in the new semantic similarity sequence as the target index identifier. Based on the target index identifier, the corresponding category label is retrieved and obtained by querying the index-tag mapping relationship pre-stored in the category standard vector library, thereby determining the material category of the application material text; When the highest value of the new semantic similarity is less than or equal to the preset low semantic similarity threshold, it is determined that the application material text cannot be reliably classified, and an abnormal classification mark is added to the application material text. Based on the application material text and all pre-stored source text data carrying the anomaly classification marker, the abnormal application materials are identified by summarizing through association identifier retrieval.

8. The intelligent classification method for application materials based on data analysis according to claim 7, characterized in that, The step of performing semantic proximity analysis on the abnormal application materials, and grouping the materials based on semantic proximity analysis results through semantic consistency evaluation to obtain the final classification result includes: Based on the abnormal application materials, the weight of each keyword is calculated using the term frequency-inverse document frequency algorithm, and a feature weight distribution matrix is ​​constructed with the application materials as rows and the keyword weights as columns; Based on the feature weight distribution matrix, the DBSCAN density clustering algorithm is used to aggregate the abnormal application materials whose mutual distance is less than the preset neighborhood radius and whose number of materials around the core point reaches the preset minimum sample number requirement, using Euclidean distance as the measure of dissimilarity between vectors, to form semantic proximity groups. Calculate the arithmetic mean of the enhanced semantic representation vector corresponding to the abnormal application materials contained in the semantic proximity group in each dimension to generate the overall semantic vector; Based on the overall semantic vector, the overall similarity between the overall semantic vector and the enhanced semantic representation vector is calculated using the cosine similarity algorithm, and the arithmetic mean of the overall similarity is obtained as the average similarity evaluation index of the semantic proximity group. Based on the average similarity evaluation index, material category labels are matched in the category standard vector library, and the material category labels are assigned to all the abnormal application materials within the semantic proximity group to determine the final classification result.

9. A data analysis-based intelligent classification system for application materials, characterized in that, include: The data acquisition module is used to obtain the text of application materials; The keyword analysis module is used to extract keywords, calculate distinction weights, and perform weighted processing based on the application material text to obtain a weighted keyword vector. This includes: segmenting the text and extracting vocabulary units using Chinese word segmentation technology to obtain word segmentation results; calculating the product of the word frequency weight and inverse document frequency weight of each vocabulary unit using a word frequency-inverse document frequency (IF-IVF) algorithm based on the word segmentation results and pre-stored corpus document frequency statistics to generate a keyword weight sequence; calculating the statistical variance of the frequency of each vocabulary unit in the application material text based on the keyword weight sequence to obtain the distinction weight of each vocabulary unit; weighted fusing the word frequency weight, the IVF weight, and the distinction weight to obtain a key feature vector; selecting key feature vectors that exceed a preset key feature threshold to form an optimized feature set; and performing normalization and vector dimension alignment based on the optimized feature set to obtain a weighted keyword vector. The semantic feature analysis module is used to map the weighted keyword vector to a high-dimensional semantic space by capturing the contextual associations between words, based on the weighted keyword vector, to obtain a multi-dimensional semantic representation vector. The semantic similarity analysis module is used to calculate the semantic similarity between vector directions one by one based on the multidimensional semantic representation vector and the preset semantic category standard vector, so as to obtain a semantic similarity sequence. The preliminary classification module is used to perform material type matching analysis based on the semantic similarity sequence and a preset semantic high similarity threshold to determine the preliminary material classification result; The preliminary output module is used to output the material type as the final classification result when the preliminary classification result of the material shows that the type match is successful. The enhanced semantic module is used to re-evaluate the distinguishing weight of keywords based on the application material text when the preliminary classification result of the material shows that the type matching fails, and to supplement and strengthen the multi-dimensional semantic representation vector based on the re-evaluated distinguishing weight to form an enhanced semantic representation vector. The anomaly screening module is used to obtain a new semantic similarity sequence by calculating the semantic similarity between the enhanced semantic representation vector and the semantic category standard vector based on the enhanced semantic representation vector, and to screen out abnormal application materials in combination with a preset semantic low similarity threshold. The final classification output module is used to perform semantic proximity analysis on the abnormal application materials, and group them according to the semantic proximity analysis results through semantic consistency evaluation to obtain the final classification results.

Images