A news text intelligent classification method and system based on deep learning

Abstract

The application relates to the technical field of natural language processing, and discloses a news text intelligent classification method and system based on deep learning, which comprises the following steps: collecting news texts and category names, normalizing the news texts and the category names to form unified corpus and a word table; counting the number of word units and the number of undirected adjacent co-occurrences to generate a Word2vec context semantic vector; generating a FastText character sub-word vector according to the number of word units and character fragments; generating a hybrid word vector by fusing the word frequency scarcity and the novelty of character fragments; superimposing a position-based function and constructing a second-order energy matrix; generating a locally reversible interaction feature through CINN and generating a hidden state through RNIN; constructing a category prototype multi-head attention feature again, inputting the category prototype multi-head attention feature into an MLP, and outputting a news category in combination with geometric energy.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method and system for intelligent classification of news text based on deep learning. Background Technology

[0002] Intelligent classification of news text is often used in scenarios such as portal information distribution, public opinion monitoring, media content archiving and industry topic retrieval. Existing methods mostly use word-level semantic representation, sub-word representation, recurrent networks, attention networks or perceptual networks to classify news texts. They can form basic classification results when dealing with news content of regular length, regular vocabulary and clear category boundaries. However, news texts are characterized by short headlines, dense body information, frequent updates of entity names, and the intermingling of synonyms. When relying solely on word-level context representation, the expressive strength of low-frequency entities, abbreviations, compound words, and newly emerging proper nouns is easily weakened. When relying solely on character or sub-word representation, the stable semantics of common words are easily interfered with by character fragments, resulting in a lack of coordination mechanism between word-level semantics and character / sub-word semantics that changes with word frequency and fragment distribution. Meanwhile, news category names are usually short, and there may be similar thematic boundaries between categories. For example, categories such as finance and industry, sports and entertainment, and technology and the Internet have a high degree of overlap in local word units. After extracting text features, the existing classification structure often finds it difficult to transform the category name itself into a category prototype that can participate in spatial matching, resulting in insufficient utilization of category-side semantic information. Furthermore, key information in news texts is not evenly distributed. Event subjects, action words, time and location, and concluding paragraphs may be distributed in different positions. There is a lack of amplitude matching between ordinary position encoding and text semantic energy. Inconsistent representation scales can also easily occur between local feature interactions and sequential state transmission, which in turn affects the stability of multi-head attention aggregation and final category evidence calculation. Therefore, there is an urgent need for a deep learning-based intelligent classification method and system for news texts to address the problems in the existing technologies, such as insufficient coordination between word-level semantics and character sub-word semantics, insufficient utilization of category prototypes, inconsistency between local interaction and sequential state scales, and insufficient stability of classification evidence. Summary of the Invention

[0003] This invention provides a method and system for intelligent classification of news text based on deep learning, which helps to solve the problems mentioned in the background art.

[0004] This invention provides the following technical solution: a deep learning-based intelligent classification method for news text, comprising: Establish a text input terminal, a category input terminal, a normalization unit, a segmentation unit, a character splitting unit, a calculation unit, and an output terminal to collect news text and category names; The news texts and category names are subjected to unified encoding, character standardization, word segmentation, and character fragment splitting to form a unified corpus and vocabulary; Count the number of lexical terms and the number of undirected neighbor co-occurrences, construct the context co-occurrence intensity matrix and the stable semantic matrix, and generate the Word2vec context semantic vector; Based on the character fragments and word frequency, the Word2vec context semantic vectors of the same character fragments are weighted and aggregated to generate FastText character sub-word vectors; A hybrid gating quantity is generated based on word frequency scarcity and character fragment novelty, and then the Word2vec context semantic vector and FastText character sub-word vector are fused to obtain a hybrid word vector; By superimposing positional basis functions on the hybrid word vectors, a hybrid representation matrix of the news text is formed. The dimensions are then split, and local contextual mean and second-order energy matrix are constructed. Based on the second-order energy matrix, scaling and translation branches of the CINN locally invertible interaction network are generated to obtain locally invertible interaction features, which are then used to generate hidden states through the RNIN sequential normalization interaction network. Category prototypes are generated from the hybrid word vectors of category name lexics. Multi-head attention features of the category prototypes are constructed, input into an MLP (Multilayer Perceptron) network, and combined with geometric energy to output news categories.

[0005] Optionally, the establishment of a text input terminal, a category input terminal, a normalization unit, a segmentation unit, a character splitting unit, a calculation unit, and an output terminal for collecting news text and category names specifically includes: Receive the news text to be categorized through the text input terminal, and receive the preset category name through the category input terminal; The news text and category names are standardized by performing unified encoding, case normalization, full-width and half-width character normalization, and punctuation boundary preservation. The news text word sequence and category name word set are generated separately by segmentation units; Each word is used to generate a set of continuous character segments by a character splitting unit, and the set of continuous character segments contains at least the corresponding word itself. The news text and category names are merged into a unified corpus by the computing unit, and a unified vocabulary is formed based on the word units that appear in the unified corpus. The news categories output by the computing unit are received at the output terminal.

[0006] Optionally, the steps of counting the number of lexical terms and the number of undirected neighbor co-occurrences, constructing a context co-occurrence strength matrix and a stable semantic matrix, and generating a Word2vec context semantic vector specifically include: Traverse the sequence of word elements in the news text and the set of word elements for category names in the unified corpus, and count the number of occurrences of each word element; Count the number of undirected co-occurrences of any two words based on their first-order adjacency relationship; The frequency of word occurrences, the frequency of word co-occurrences, the total number of co-occurrences, and the size of the vocabulary are smoothed and normalized to generate a context co-occurrence intensity matrix. A stable semantic matrix is ​​obtained by adding a stabilization term determined by the total number of co-occurrences and the vocabulary size to the context co-occurrence intensity matrix. Perform singular value decomposition on the stable semantic matrix and determine the retained dimension based on the direction of the singular values ​​that is not lower than the mean of the singular values; The retained directions are sorted in descending order of singular values. When there are equal singular values, the standard basis vectors are projected into a subspace according to the unified vocabulary index order, skipping zero projection. Non-zero projections are orthogonalized and the vector directions are fixed. Extract the Word2vec context semantic vector for each word.

[0007] Optionally, the step of weighted aggregation of Word2vec context semantic vectors of the same character segment based on character fragments and word frequency to generate FastText character sub-word vectors specifically includes: Read the set of consecutive character segments generated by the character splitting unit; Retrieve lexicons containing the same character segment from a unified thesaurus; Treat the lexical units containing the same character segment as aggregate objects; Using the frequency of occurrence of each aggregated object in the unified corpus as the weighting criterion, a weighted average is performed on the Word2vec context semantic vector of each aggregated object to generate the FastText character fragment vector of the corresponding character fragment; The FastText character fragment vectors contained in the target word are averaged to generate FastText character sub-word vectors for the target word. The FastText character sub-word vectors are kept to have the same dimension as the Word2vec context semantic vectors of the target words.

[0008] Optionally, the step of generating a hybrid gating quantity based on word frequency scarcity and character fragment novelty, and fusing Word2vec contextual semantic vectors and FastText character sub-word vectors to obtain a hybrid word vector specifically includes: Extract the highest word frequency and the median word frequency based on the occurrence frequency of each word element in the unified vocabulary; The character segments contained in words that appear at least once the median frequency are merged into a common character segment set; Term frequency scarcity is generated based on the ratio between the number of occurrences of the target term and the highest term frequency. The novelty of a character fragment is generated based on the overlap ratio between the target word metacharacter fragment set and the common character fragment set. The gating base quantity is obtained by summing the word frequency scarcity and the character fragment novelty. Using the basic gating quantity as the numerator of the occupancy ratio and the sum of the basic gating quantity and the fixed dimensionless normalized benchmark quantity as the denominator of the occupancy ratio, the occupancy ratio of FastText character sub-word vectors is extracted to obtain the hybrid gating quantity. The proportion of FastText character sub-word vectors is controlled by a hybrid gating variable, and the remaining proportion of Word2vec context semantic vectors is used for complementary fusion to generate hybrid word vectors of target words.

[0009] Optionally, the step of superimposing positional basis functions on the hybrid word vectors to form a hybrid representation matrix of the news text, splitting the dimensions, and constructing local contextual mean and second-order energy matrices specifically includes: Read the mixed word vector of each word element according to the word element sequence of the news text; Generate a sinusoidal positional basis function based on word position, news text length, and vector dimension; Calculate the position amplitude matching quantity based on the overall energy of the mixed word vectors and the overall energy of the position basis functions within the current news text. The positional basis functions after positional amplitude matching are superimposed on the hybrid word vectors of the corresponding word units to generate a hybrid representation matrix of news text; The news text mixed representation matrix is ​​split into a first subspace matrix and a second subspace matrix according to the dimensional order; Based on the set of first-order adjacent positions of each word, a local average is performed on the first subspace matrix to generate a local context average. Perform transpose product and length normalization on the mixed representation matrix of news text to generate a second-order energy matrix.

[0010] Optionally, the step of generating scaling and translation branches of the CINN locally invertible interaction network based on the second-order energy matrix to obtain locally invertible interaction features, and generating hidden states through the RNIN sequential normalization interaction network, specifically includes: From the second-order energy matrix, extract the autocorrelation block of the first subspace and the cross-correlation block from the first subspace to the second subspace; The scaling branch matrix of the CINN locally invertible interactive network is generated based on the autocorrelation blocks of the first subspace, the anti-zero smoothing amount determined by the total co-occurrence and vocabulary size, and the cross-correlation blocks. Based on the principal value symmetric inverse square root and cross-correlation blocks of the first subspace autocorrelation block, the translation branch matrix of the CINN locally invertible interactive network is generated. Affine coupling is performed on the second subspace features by local context mean, scaling branch matrix and translation branch matrix, and then concatenated with the first subspace features to obtain CINN locally invertible interactive features; Based on the locally reversible interaction features of CINN, construct the second-order energy matrix and whitening projection matrix of the interaction features; Based on the cosine consistency of the local reversible interaction features of adjacent CINNs, the sequential memory gating quantity of RNIN is generated. The current CINN local reversible interaction features, the previous hidden state, and the RNIN sequential memory gating input are processed by layer normalization operations determined by the mean and variance to generate the RNIN hidden state at the current position.

[0011] Optionally, the step of generating category prototypes based on the mixed word vectors of category name lexics and constructing multi-head attention features for category prototypes specifically includes: Each category name is segmented using the same word segmentation method as the news text; Read the mixed word vectors of the terms contained in the category name, and perform averaging on the mixed word vectors to generate the category prototype; Read the whitening projection matrix generated by the RNIN sequential normalized interactive network; The category prototypes are input into the whitening projection matrix and the layer normalization operation to generate the order space category prototypes. Construct a category prototype covariance matrix based on the order space category prototypes corresponding to all news texts and all categories; The number of attention heads is determined based on the algebraic rank of the category prototype covariance matrix; The projection matrix of the corresponding attention head is generated based on the non-zero feature direction. When the category prototype covariance matrix degenerates, the dimension-normalized unit projection matrix is ​​used. Calculate the category-related score based on the RNIN hidden state, projection matrix, and order space category prototype; Centralized index normalization is applied to category-related scores to obtain attention weights; The RNIN hidden states are weighted and aggregated according to the attention weights to obtain the multi-head attention features of the category prototype.

[0012] Optionally, the step of inputting the news category into the MLP (Multilayer Perceptron) and combining it with geometric energy specifically includes: The local reversible interaction features of CINN are averaged along the word positions to obtain the local interaction mean. The RNIN hidden states are averaged along the word positions to obtain the sequential state mean. The local interaction mean, the sequential state mean, the category prototype multi-head attention feature, and the sequential space category prototype are concatenated into a news combination feature; The sequential space category prototype is extended in the same dimension according to the composition structure of news combination features to form category anchoring features; Generate a mean vector, covariance matrix, and whitening weight matrix based on the combined features of all news items; News combination features and category anchoring features are mapped to the same whitening space, and category evidence is generated by error function-type nonlinear activation operation of MLP multilayer perceptron. Standardize the amount of evidence for each category within the same news text; Normalized geometric energy is generated based on the squared norm distance between the multi-head attention features of the category prototype and the category prototype in the order space. The difference between the standardized category evidence quantity and the normalized geometric energy is calculated to obtain the final classification score; Output the category with the highest final classification score as the news category. When multiple categories have the same final classification score, output the category with the smallest index.

[0013] A system for implementing the deep learning-based intelligent news text classification method includes: Input preprocessing module: used to collect news text and category names, and generate a unified corpus and vocabulary; Semantic representation module: used to count word co-occurrence relationships and generate context semantic vectors; Sub-word representation module: used to aggregate character fragment information and generate character sub-word vectors; Hybrid representation module: used to generate hybrid word vectors based on context semantic vectors and character sub-word vectors; Sequence matrix module: used to perform positional encoding on mixed word vectors and generate a second-order energy matrix; CINN Interactive Module: Used to perform reversible interactive transformations on the second-order energy matrix and output locally reversible interactive features; RNIN interaction module: used to generate hidden states based on local reversible interaction features; Attention aggregation module: used to generate category prototypes based on hidden states and aggregate multi-head attention features based on category prototypes; Classification output module: Used to generate category evidence based on multi-head attention features and output news categories.

[0014] The present invention has the following beneficial effects: 1. By setting up a text input terminal, a category input terminal, a normalization unit, a segmentation unit, a character splitting unit, a calculation unit, and an output terminal, and by performing unified encoding, case normalization, full-width and half-width normalization, punctuation boundary preservation, word segmentation, and character fragment splitting on news text and category names, text objects from the main text title, category label, and entity words can enter the same processing scale, reducing the disturbance of encoding differences, character morphology differences, and word segmentation boundary differences on subsequent semantic representation. At the same time, by establishing the basis for subsequent matrix row and column positioning through a unified corpus and a unified vocabulary, the information on the news text side and the information on the category name side can participate in the calculation within the same index system; 2. By statistically analyzing the occurrence frequency of word elements and the number of first-order adjacent undirected co-occurrences in a unified corpus, a context co-occurrence strength matrix and a stabilized semantic matrix are constructed. This allows the local co-occurrence relationships between adjacent word elements in news texts to be transformed into a decomposable semantic space structure. Non-negative truncation reduces the interference of weak correlations or noisy co-occurrences on the direction of word vectors. Stabilization terms alleviate computational instability caused by zero co-occurrence and matrix degradation. Singular value mean filtering and repeated root deterministic processing enable Word2vec context semantic vectors to maintain stable direction selection and dimensional boundaries, thereby enhancing the consistency of word-level contextual expression. 3. By reading a set of continuous character fragments and retrieving lexical units containing the same character fragment from a unified vocabulary, the corresponding Word2vec context semantic vectors are aggregated based on the frequency of lexical unit occurrences as a weighting criterion. This enables character fragments in low-frequency words, new compound words, abbreviations, and proper entities to obtain semantic compensation from relevant lexical units, avoiding the problem of weak expression of low-frequency news entities when relying solely on the frequency of whole word occurrences. At the same time, by performing average processing on all character fragment vectors of the target lexical unit, the FastText character sub-word vectors and word-level semantic vectors maintain the same dimension relationship, providing sub-word-side expressions that can be directly fused for subsequent hybrid representations. 4. By generating a hybrid gating quantity based on the highest word frequency, median word frequency, common character fragment set, word frequency scarcity, and character fragment novelty, and using this gating quantity to adjust the proportion of FastText character sub-word vectors and Word2vec context semantic vectors, high-frequency common words are more biased towards stable context semantic expression, and low-frequency words and novel character fragments are more fully incorporated into sub-word morphological information. This alleviates the problem that single word vectors are difficult to handle new words and that single sub-word vectors are prone to interfering with the semantics of common words. Therefore, the hybrid word vector can maintain a more reasonable semantic allocation in scenarios of news entity changes and topic word reuse. 5. By generating a sinusoidal positional basis function based on word position, news text length, and vector dimension, and calculating the position amplitude matching amount based on the overall energy of the hybrid word vector and the stable energy of the positional basis function, the positional information can maintain scale coordination with the semantic energy of the text when superimposed on the hybrid word vector, reducing the representational offset caused by excessively strong or weak positional terms. At the same time, the first and second subspace matrices are formed by dimensional splitting, and the local context mean is generated based on the first-order adjacent position set, so that the local phrase structure in the news text can enter the subsequent interaction calculation. The second-order energy matrix provides a stable statistical basis for local reversible interactions. 6. By extracting the autocorrelation blocks of the first subspace and the cross-correlation blocks from the first subspace to the second subspace from the second-order energy matrix, and generating the CINN scaling branch matrix and translation branch matrix, the features of the second subspace can complete affine coupling under the constraint of local context mean, while preserving the features of the first subspace and enhancing the local structural expression. Subsequently, based on the local reversible interaction features of CINN, the RNIN whitening projection matrix and sequential memory gating quantity are generated, so that the consistency of adjacent positions can adjust the degree of fusion between historical states and candidate states, and reduce the scale perturbation caused by local mutations to the sequential hidden state. 7. By averaging the mixed word vectors of category name lexical units to generate category prototypes, and then forming order space category prototypes through RNIN whitening projection matrix and layer normalization, short category names can be transformed into category-side semantic objects that can be compared with the hidden state of news. The number of attention heads is then determined by the algebraic rank of the category prototype covariance matrix, and a projection matrix is ​​generated based on the non-zero feature direction, so that the difference direction between different categories participates in multi-head attention aggregation. The centralization index normalization process further reduces the dominance of large value scores on weight distribution, thereby enabling category name semantics to participate in the key position screening of news text. 8. By performing positional mean processing on the local reversible interaction features of CINN and the hidden state of RNIN respectively, and concatenating them with the multi-head attention features of category prototypes and the category prototypes in the order space to form news combination features, the local structure, order state, category attention information and category prototype information are jointly entered into the MLP evidence compression process. The whitening weight matrix and error function type nonlinear activation make the category evidence quantity comparable in the same whitening space. Furthermore, the geometric deviation between the multi-head attention features of category prototypes and the category prototypes in the order space is combined to generate the final classification score, so that the classification output is simultaneously constrained by the evidence strength and the consistency of the category space. 9. This solution targets fine-grained text classification scenarios in news distribution, public opinion monitoring, and content archiving. Addressing the classification instability caused by short titles, low-frequency entities, similar category boundaries, and uneven semantic location distribution, it constructs a hybrid representation link based on Word2vec contextual semantic vectors and FastText character sub-word vectors. It collaboratively combines CINN locally reversible interactions, RNIN sequential normalization interactions, category prototype multi-head attention, and MLP evidence compression. Specifically, the Word2vec channel provides the semantic skeleton of word context, while the FastText channel supplements the expression of low-frequency and novel words by character fragments. Hybrid gating enables the two to dynamically complement each other based on word frequency scarcity and character fragment novelty. CINN decomposes the subspace... After segmentation, the local context mean is transformed into a reversible affine interaction, preventing the local structure of the text from being unidirectionally compressed. RNIN then utilizes the consistency of adjacent interaction features to generate sequential hidden states, mitigating the interference of local mutations on the global state. Multi-head attention of category prototypes introduces the semantics of category names into the text aggregation process, allowing short category labels to participate in feature selection. The difference between MLP evidence quantity and geometric energy further constrains the spatial deviation between category evidence and category prototypes. This forms a collaborative discrimination mechanism among word-level semantics, sub-word morphology, local interactions, sequential states, category prototypes, and classification evidence, enabling unified processing of low-frequency entity recognition, nearest neighbor category differentiation, and stable evidence output problems that are difficult to handle simultaneously by individual word vectors, individual sub-word models, individual attention, or individual perceptual networks. Attached Figure Description

[0015] Figure 1 This is a flowchart of the intelligent news text classification method based on deep learning according to the present invention.

[0016] Figure 2 This is a schematic diagram illustrating the generation of a hybrid representation using Word2vec and FastText according to the present invention.

[0017] Figure 3 Hybrid gated surface plot driven by word frequency scarcity and character fragment novelty.

[0018] Figure 4 This is a heat map of the mixed representation matrix and second-order energy matrix of news text.

[0019] Figure 5 A schematic diagram of the category prototype, news combination features, and geometric energy classification space.

[0020] Figure 6 The login interface of the system of this invention.

[0021] Figure 7 The main interface diagram of the system of this invention.

[0022] Figure 8The interface diagram of the input preprocessing and corpus construction of the system of the present invention.

[0023] Figure 9 The interface diagram for generating context semantic vectors using the Word2vec system of this invention.

[0024] Figure 10 The interface diagram for generating character sub-word vectors using the FastText system of this invention.

[0025] Figure 11 The interface diagram of the dynamic fusion of hybrid word vectors in this invention system.

[0026] Figure 12 The present invention provides a system sequence representation and a locally reversible interactive interface diagram.

[0027] Figure 13 The system category prototype attention and classification output interface diagram of this invention. Detailed Implementation

[0028] 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.

[0029] Example 1: A Deep Learning-Based Intelligent Classification Method for News Text, referring to... Figure 1 As shown, it includes the following steps: Step 1: Establish an input processing environment and form a unified corpus and vocabulary; Set up a text input terminal, a category input terminal, a normalization unit, a segmentation unit, a character splitting unit, a calculation unit, and an output terminal; The text input terminal receives the news text to be categorized; the category input terminal receives the preset category name; the normalization unit performs unified encoding, case normalization, full-width and half-width normalization, and punctuation boundary preservation on the news text to be categorized and the preset category name; The segmentation unit performs word segmentation on the news text to generate a word sequence of the news text; the segmentation unit performs word segmentation on the category name using the same rules to generate a set of category name words. The character splitting unit generates a set of consecutive character segments for each word, and ensures that the set of consecutive character segments for each word contains at least the word itself; The calculation unit merges the news text to be classified with the preset category name to form a unified corpus; the calculation unit traverses all the word elements that appear in the unified corpus and builds a unified vocabulary and corresponding word element index according to the order of first appearance; the word element index is used for subsequent matrix row and column positioning; In this embodiment, the Word2vec context semantic vector, FastText character sub-word vector, CINN locally reversible interactive network, RNIN sequential normalized interactive network, and MLP multilayer perceptron network all refer to the domain-specific self-calibrated computational structure constructed from a unified corpus, category names, lexical statistics, character segment statistics, and prior calculation results. The model names mentioned above are used to define the data processing methods and functional locations of each computational structure; they do not indicate the use of external pre-trained word vectors, external deep learning models, external labeled training sets, or external model parameters. In this embodiment, the word vectors, category prototype vectors, hidden state vectors, and combined feature vectors are all in row vector form; the linear transformation matrices are all applied to the row vectors using right multiplication. By setting up a text input terminal, a category input terminal, a normalization unit, a segmentation unit, a character splitting unit, a calculation unit, and an output terminal, and by performing unified encoding, case normalization, full-width and half-width normalization, punctuation boundary preservation, word segmentation, and character fragment splitting on news text and category names, text objects from the main text title, category label, and entity words can enter the same processing scale, reducing the disturbance of encoding differences, character morphology differences, and word segmentation boundary differences on subsequent semantic representation. At the same time, by establishing the basis for subsequent matrix row and column positioning through a unified corpus and a unified vocabulary, the information on the news text side and the information on the category name side can participate in the calculation within the same index system. Step 2: Construct the Word2vec context semantic vector; Reference Figure 2 This step, along with subsequent steps three and four, together form the Word2vec and FastText hybrid representation generation process. The Word2vec context semantic vector is determined by the word co-occurrence statistics and the stabilized semantic matrix within the unified corpus. The FastText character sub-word vector is determined by the character fragment set, word occurrence frequency, and Word2vec context semantic vector. The hybrid gating quantity is determined by word frequency scarcity and character fragment novelty, and is used to adjust the ratio of Word2vec context semantic vector to FastText character sub-word vector. In this step, the Word2vec context semantic vector is determined by the word co-occurrence statistics and stable semantic matrix within the unified corpus, without calling external Word2vec pre-trained word vectors; The computational unit traverses the sequence of word elements in the news text and the set of word elements for category names in the unified corpus, and counts the number of times each word element appears in the unified corpus; The calculation unit counts the number of undirected co-occurrences of any two words as adjacent words in a unified corpus, based on the first-order adjacency relationship; the left and right order of the same pair of adjacent words is not distinguished. The calculation unit sums up all undirected co-occurrence counts to obtain the total co-occurrence count; the reciprocal of the sum of the total co-occurrence count and the square of the unified vocabulary size is taken to obtain the zero-prevention smoothing amount; For any two lexical units in the unified vocabulary, the computing unit first adds the number of undirected co-occurrences of the two units to the zero-prevention smoothing amount to obtain the first smooth co-occurrence amount; then it adds the total co-occurrence amount to the result obtained by multiplying the zero-prevention smoothing amount by the square of the unified vocabulary size to obtain the second smooth co-occurrence amount. The calculation unit multiplies the first smooth co-occurrence quantity and the second smooth co-occurrence quantity to obtain the co-occurrence intensity numerator; it adds the occurrence counts of the two words to the anti-zero smoothing quantity multiplied by the unified vocabulary size, and then multiplies the two to obtain the co-occurrence intensity denominator. The calculation unit divides the co-occurrence intensity numerator by the co-occurrence intensity denominator and takes the natural logarithm, and performs non-negative truncation on the natural logarithm result; when the natural logarithm result is less than zero, it is taken as zero, and when the natural logarithm result is not less than zero, the result is retained; thus, the context co-occurrence intensity between two tokens is obtained. The calculation unit arranges the context co-occurrence intensities among all lexical units according to a unified vocabulary index and generates a context co-occurrence intensity matrix. The computational unit adds a stabilization term to the context co-occurrence intensity matrix to obtain a stable semantic matrix; the stabilization term is formed by adding a zero-prevention smoothing amount to the main diagonal of the context co-occurrence intensity matrix, and the stabilization term is not added to the non-main diagonal positions; The computational unit performs singular value decomposition on the stabilized semantic matrix, obtaining singular values, left singular vectors, and singular value diagonal matrices arranged in descending numerical order; The computing unit calculates the arithmetic mean of all singular values, selects the singular value directions whose values ​​are not lower than the average of the singular values ​​as retained directions, counts the number of retained directions, and obtains the initial retained dimension. If the initial retained dimension is less than 1, then the initial retained dimension is fixed to 1; if the initial retained dimension is less than 2, then zero columns are padded to make the final vector dimension no less than 2, so as to meet the subsequent CINN dimension splitting requirements. When equal singular values ​​exist, the computation unit performs deterministic basis vector generation processing on the corresponding singular subspace; specifically, it reads the standard basis vectors sequentially according to the unified vocabulary index order and projects the standard basis vectors onto the singular subspace; if the projection result is a zero vector, the standard basis vector is skipped; if the projection result is a non-zero vector, the projection result is orthogonalized. The computational unit continues to process in the above order until the required number of orthogonal basis vectors are obtained; the direction of each orthogonal basis vector is fixed, making its first non-zero component positive; The computation unit generates a Word2vec context semantic vector matrix based on the retained left singular vector and the corresponding square root of the singular value. If the final vector dimension is greater than the initial retained dimension, zero columns are added to the Word2vec context semantic vector matrix to make the matrix dimension consistent with the final vector dimension. For each word in the unified vocabulary, the computing unit reads the corresponding row of the word in the Word2vec context semantic vector matrix, divides the row vector by the sum of its L2 norm and zero-prevention smoothing factor, and obtains the Word2vec context semantic vector of the word. In this step, the number of word occurrences and the number of undirected co-occurrences are obtained from the unified corpus statistics; the zero-smoothing amount is determined by the total co-occurrence amount and the vocabulary size; the Word2vec context semantic vector is the output of this step and is used for subsequent FastText character sub-word vector and hybrid word vector calculations; By statistically analyzing the occurrence frequency of word elements and the number of first-order adjacent undirected co-occurrences in a unified corpus, and constructing a context co-occurrence strength matrix and a stabilized semantic matrix, the local co-occurrence relationships between adjacent word elements in news texts can be transformed into a decomposable semantic space structure. Non-negative truncation reduces the interference of weak correlations or noisy co-occurrences on the direction of word vectors, stabilization terms alleviate computational instability caused by zero co-occurrence and matrix degradation, and singular value mean filtering and repeated root deterministic processing enable Word2vec context semantic vectors to maintain stable direction selection and dimensional boundaries, thereby enhancing the consistency of word-level contextual expression. Step 3: Generate FastText character sub-word vectors; In this step, the FastText character sub-word vectors are determined by the character fragment set, the number of word occurrences, and the Word2vec context semantic vector, without calling the external FastText pre-trained sub-word model; The computing unit reads the set of continuous character segments generated by the character splitting unit; for each character segment, it retrieves all words containing that character segment from the unified vocabulary and uses these words as the aggregation object of that character segment; For each aggregated object, the computing unit reads its occurrence count in the unified corpus and the Word2vec context semantic vector generated in step two; The calculation unit multiplies the occurrence count of each aggregated object with its Word2vec context semantic vector to obtain the weighted semantic vector of that aggregated object; The calculation unit sums all weighted semantic vectors corresponding to the same character segment to obtain the total weighted semantic vector of the character segment; it sums the occurrence counts of all aggregated objects corresponding to the same character segment and adds them to the zero-prevention smoothing amount to obtain the normalized denominator of the character segment. The calculation unit divides the weighted semantic vector of the character fragment by the normalized denominator of the character fragment to generate the FastText character fragment vector of that character fragment; For a target word, the computing unit reads the FastText character fragment vectors corresponding to all character fragments in its continuous character fragment set; after summing these FastText character fragment vectors, it divides them by the number of character fragments contained in the target word to generate the FastText character sub-word vector of the target word; Since the FastText character fragment vector is obtained by weighted aggregation of the Word2vec context semantic vector, the FastText character sub-word vector and the Word2vec context semantic vector maintain the same dimension; This step outputs FastText character sub-word vectors, which are then used for subsequent mixed word vector calculations; the character fragment aggregation weights in this step are determined by the frequency of word occurrences in the unified corpus. By reading a continuous set of character fragments and retrieving lexical units containing the same character fragment from a unified vocabulary, the corresponding Word2vec context semantic vectors are aggregated based on the frequency of lexical unit occurrences as a weighting criterion. This enables character fragments in low-frequency words, new compound words, abbreviations, and proper entities to obtain semantic compensation from relevant lexical units, avoiding the problem of weak expression of low-frequency news entities when relying solely on the frequency of whole words. At the same time, by performing average processing on all character fragment vectors of the target lexical unit, the FastText character sub-word vectors and word-level semantic vectors maintain the same dimension relationship, providing sub-word-side expressions that can be directly fused for subsequent hybrid representations. Step 4: Generate hybrid word vectors combining Word2vec and FastText; The calculation unit reads the occurrence count of each word element in the unified vocabulary, extracts the highest word frequency, and calculates the median word frequency. When the size of the unified vocabulary is odd, the median frequency of the middle word after frequency sorting is used as the median frequency; when the size of the unified vocabulary is even, the arithmetic mean of the two middle word frequencies after frequency sorting is used as the median frequency. The calculation unit takes lexical units that appear at least once the median frequency as a common lexical unit set, and merges the character segments contained in these common lexical units to form a common character segment set; For the target word, the calculation unit divides the number of times the target word appears by the highest word frequency to obtain the word frequency percentage; and subtracts the word frequency percentage from 1 to obtain the word frequency scarcity. The closer the frequency of the target word is to the highest word frequency, the lower the word frequency scarcity; the lower the frequency of the target word, the higher the word frequency scarcity. The computing unit reads the set of character fragments of the target word and calculates the number of intersection fragments between the set of character fragments of the target word and the set of common character fragments; The calculation unit divides the number of intersection fragments by the sum of the number of target word metacharacter fragments and the zero-smoothing amount to obtain the common proportion of character fragments; subtracting the common proportion of character fragments from 1 yields the novelty of the character fragments. The higher the overlap ratio, the lower the novelty of the character fragment; the lower the overlap ratio, the higher the novelty of the character fragment. Reference Figure 3 The scarcity of word frequency and the novelty of character fragments jointly determine the proportion of FastText character sub-word vectors in the mixed word vectors. Specifically, the higher the scarcity of word frequency and the higher the novelty of character fragments, the higher the proportion of FastText character sub-word vectors corresponding to the mixed gate. Conversely, the lower the scarcity of word frequency and the lower the novelty of character fragments, the lower the proportion of FastText character sub-word vectors corresponding to the mixed gate. The calculation unit sums the word frequency scarcity and the character fragment novelty to obtain the gating base quantity; the gating base quantity is used as the numerator of the occupancy ratio; the sum of the gating base quantity and the fixed dimensionless normalized benchmark quantity is used as the denominator of the occupancy ratio. The calculation unit divides the numerator of the occupancy ratio by the denominator of the occupancy ratio to obtain the occupancy ratio of the FastText character sub-word vector, and uses this occupancy ratio as the hybrid gating quantity; it subtracts the hybrid gating quantity from 1 to obtain the occupancy ratio of the Word2vec context semantic vector. The computing unit multiplies the Word2vec context semantic vector by its occupancy ratio, multiplies the FastText character sub-word vector by the hybrid gating quantity, and then adds the two together to generate the hybrid word vector of the target word. The fixed dimensionless normalized reference quantity is the fixed unit reference quantity used for gate proportional closure in this embodiment, and its value is fixed at 1; In this embodiment, the fixed dimensionless normalization benchmark is only used to keep the proportion of FastText character sub-word vectors within a finite closed range. If the normalization benchmark is too large, the proportion of FastText character sub-word vectors decreases overall, and the hybrid word vectors are closer to the Word2vec context semantic vectors. If the normalization benchmark is too small, the proportion of FastText character sub-word vectors increases overall, and the influence of low-frequency words and novel character fragments on the hybrid word vectors is enhanced. This step outputs hybrid word vectors, which can then be used to construct the hybrid representation matrix of the news text. By generating a hybrid gating quantity based on the highest word frequency, median word frequency, common character fragment set, word frequency scarcity, and character fragment novelty, and using this gating quantity to adjust the proportion of FastText character sub-word vectors and Word2vec context semantic vectors, high-frequency common words are more biased towards stable context semantic expression, and low-frequency words and novel character fragments are more fully incorporated into sub-word morphological information. This alleviates the problem that single word vectors are difficult to handle new words and that single sub-word vectors are prone to interfering with the semantics of common words. Therefore, the hybrid word vector can maintain a more reasonable semantic allocation in scenarios of news entity changes and topic word reuse. Step 5: Generate the mixed representation matrix of the news text, the local context mean, and the second-order energy matrix; Reference Figure 4 This step arranges the mixed word vectors corresponding to the word sequence of the news text into a mixed representation matrix of the news text, and forms a matrix representation with positional information by superimposing positional basis functions; then, the mixed representation matrix of the news text is divided into a first subspace and a second subspace according to the vector dimension, and a second-order energy matrix is ​​formed based on the transpose product and length normalization of the mixed representation matrix of the news text. The computing unit reads the mixed word vector of each word element according to the word element sequence of the news text; For the position from the first word to the last word, and for the first vector dimension to the last vector dimension, the computation unit generates a sinusoidal positional basis function; The independent variable of the sinusoidal positional basis function is: twice the circumference constant, the product of the word position number and the vector dimension number, divided by the product of the news text length plus 1 and the vector dimension plus 1; the calculation unit takes the sine value of this independent variable to obtain the positional basis function value of the corresponding position and the corresponding dimension. The computing unit calculates the sum of the squared values ​​of all mixed word vector components in the current news text to obtain the overall energy of the mixed word vectors. The calculation unit calculates the sum of the squared values ​​of all position basis function values ​​in the current news text and adds it to the zero-prevention smoothing amount to obtain the position basis function stabilization energy; The computing unit divides the overall energy of the mixed word vector by the stable energy of the position basis function, and then takes the square root of the ratio to obtain the position amplitude matching quantity. The calculation unit multiplies the position amplitude matching value by the position basis function value of the corresponding position and the corresponding dimension, and adds the result to the same dimension component of the corresponding word vector to generate the news text hybrid representation matrix; The computational unit divides the final vector dimension by 2 and rounds down to obtain the first subspace dimension; it then subtracts the first subspace dimension from the final vector dimension to obtain the second subspace dimension. The calculation unit divides the first dimension of the news text hybrid representation matrix into the first subspace matrix according to the dimensional order, and divides the remaining dimensions into the second subspace matrix. For each word element position, the computing unit constructs a set of first-order adjacent positions for that position; the set of first-order adjacent positions includes the position itself, the previous position, and the next position, and removes positions that exceed the boundaries of the news text. The computing unit reads the row vectors belonging to the first-order adjacent position set in the first subspace matrix, and performs an arithmetic mean on these row vectors to obtain the local context mean of the word position; The computing unit performs transpose multiplication on the mixed representation matrix of the news text, and then divides it by the length of the news text to generate a second-order energy matrix; This step outputs the news text mixed representation matrix, the first subspace matrix, the second subspace matrix, the local context mean, and the second-order energy matrix; these data are used for subsequent calculations by the CINN locally invertible interactive network. By generating a sinusoidal positional basis function based on word position, news text length, and vector dimension, and calculating the position amplitude matching amount based on the overall energy of the hybrid word vector and the stable energy of the positional basis function, the positional information can maintain scale coordination with the semantic energy of the text when superimposed on the hybrid word vector, reducing the representational offset caused by excessively strong or weak positional terms. At the same time, the first and second subspace matrices are formed by dimensional splitting, and the local context mean is generated based on the first-order adjacent position set, so that the local phrase structure in the news text can enter the subsequent interaction calculation. The second-order energy matrix provides a stable statistical basis for local reversible interactions. Step 6: Generate CINN locally reversible interactive features and RNIN hidden states; The locally invertible interactive CINN network in this step is determined by the second-order energy matrix, the autocorrelation block of the first subspace, the cross-correlation block from the first subspace to the second subspace, and the local context mean, without calling external CINN model parameters; In this step, the RNIN sequential normalized interaction network is determined by the CINN local invertible interaction features, the second-order energy matrix of the interaction features, the RNIN whitening projection matrix, and the sequential memory gating quantity, without calling external RNIN model parameters. The computational unit extracts autocorrelation blocks between the first subspace and the second subspace from the second-order energy matrix; The computational unit adds the result of multiplying the first subspace autocorrelation block by the anti-zero smoothing amount and the first subspace identity matrix to obtain the first stable block; The computational unit inverts the first stable block and multiplies the inversion result by the cross-correlation block from the first subspace to the second subspace to generate the scaling branch matrix of the CINN locally invertible interactive network. The computing unit calculates the principal value symmetric inverse square root of the first stable block, and then multiplies the result of the principal value symmetric inverse square root by the cross-correlation block from the first subspace to the second subspace to generate the translation branch matrix of the CINN locally invertible interactive network. For each word position in the news text, the computing unit reads the first subspace feature, the second subspace feature, and the local context mean of that position; The computational unit right-multiplies the local context mean by the scaling branch matrix to obtain the scaling driving vector; it first performs hyperbolic tangent operation on the scaling driving vector, and then performs natural exponential operation to obtain a positive scaling vector. The computational unit performs element-wise multiplication of the second subspace features with the positive scaling vector to obtain the scaled second subspace features; The computational unit multiplies the local context mean by the translation branch matrix to obtain the translation vector; it then adds the scaled second subspace features to the translation vector to obtain the affine-coupled second subspace features. The computational unit concatenates the features of the first subspace with the features of the second subspace after affine coupling to obtain the CINN locally reversible interactive features at that location; since the result of the natural exponent operation is always positive, the transformation of the second subspace has the condition for reversible recovery; The computing unit constructs an interaction feature matrix based on the local invertible interaction features of CINN at all locations; the interaction feature matrix is ​​then transposed and multiplied, and then divided by the length of the news text to obtain the second-order energy matrix of the interaction features. The computational unit adds the result of multiplying the second-order energy matrix of the interaction features by the anti-zero smoothing amount and the final vector dimension identity matrix, and performs principal value symmetric inverse square root processing on the resulting matrix to obtain the RNIN whitening projection matrix. For each word position, the computation unit reads the local reversible interaction features of the CINN at the current position and the local reversible interaction features of the CINN at the previous position; the local reversible interaction features of the CINN before the first position are taken as zero vectors; The calculation unit takes the inner product of the current position feature and the previous position feature to obtain the numerator of the adjacent feature similarity; it multiplies the L2 norm of the current position feature with the L2 norm of the previous position feature and adds it to the anti-zero smoothing amount to obtain the denominator of the adjacent feature similarity. The computational unit divides the numerator of adjacent feature similarity by the denominator of adjacent feature similarity to obtain cosine consistency; it adds one to the cosine consistency and then multiplies it by one-half to obtain the RNIN sequential memory gating quantity. The computation unit reads the RNIN hidden state from the previous position; the RNIN hidden states before the first position are taken as zero vectors. The computational unit right-multiplies the current position's CINN local reversible interactive features by the RNIN whitening projection matrix to obtain the current whitening features; The computational unit performs element-wise multiplication of the previous RNIN hidden state with the current CINN local reversible interactive feature, and then multiplies the result by the RNIN whitening projection matrix to obtain the interactive whitening feature. The calculation unit adds the current whitening feature to the interactive whitening feature and then performs a hyperbolic tangent operation to obtain the candidate state; The computation unit multiplies the previous RNIN hidden state by the RNIN sequential memory gate, multiplies the candidate state by 1 and subtracts the RNIN sequential memory gate from the result, and then adds the two together to obtain the hidden state to be normalized. The computation unit performs layer normalization on the hidden state to be normalized; the layer normalization includes: calculating the mean of each dimension component of the hidden state to be normalized; subtracting the full-dimensional row vector corresponding to the mean from the hidden state to be normalized; calculating the variance of each dimension component of the hidden state to be normalized relative to the mean; adding the variance to the anti-zero smoothing amount and taking the square root; dividing the result of removing the mean by the square root to obtain the RNIN hidden state at the current position; Among them, the principal value symmetric inverse square root is a deterministic matrix function; the central mapping coefficient of cosine consistency is taken as half, which is used to map cosine consistency from the signed interval to the non-negative gated interval; if the mapping coefficient is too large, the proportion of historical hidden states changes too slowly; if the mapping coefficient is too small, the proportion of current local invertible interaction features changes too strongly. This step outputs the CINN local reversible interaction features, RNIN whitening projection matrix, RNIN sequential memory gating amount, and RNIN hidden state, which can be used for subsequent category prototype multi-head attention features and MLP classification calculations; By extracting the autocorrelation blocks of the first subspace and the cross-correlation blocks from the first subspace to the second subspace from the second-order energy matrix, and generating the CINN scaling branch matrix and translation branch matrix, the features of the second subspace can complete affine coupling under the constraint of local context mean, while preserving the features of the first subspace and enhancing the local structural expression. Subsequently, based on the local reversible interaction features of CINN, the RNIN whitening projection matrix and sequential memory gating quantity are generated, so that the consistency of adjacent positions can adjust the degree of fusion between historical states and candidate states, and reduce the scale perturbation caused by local mutations to the sequential hidden state. Step 7: Generate category prototype multi-head attention features; The calculation unit uses the same word segmentation method as the news text for each category name; The computing unit reads the mixed word vectors of the words contained in the category name, sums these mixed word vectors and divides them by the number of words contained in the category name to generate the category prototype; The computational unit reads the RNIN whitening projection matrix generated in step six; after right-multiplying the category prototype by the RNIN whitening projection matrix, it performs layer normalization to generate the order space category prototype; The calculation unit summarizes all news texts and all categories and their corresponding order space category prototypes, and calculates the mean of these order space category prototypes; For each order space category prototype, the computational unit calculates its deviation row vector relative to the mean; the transpose column vector of the deviation row vector is multiplied by the deviation row vector to obtain the outer product matrix corresponding to that order space category prototype. The computational unit sums all the outer product matrices and divides them by the product of the number of news texts and the number of categories to construct the category prototype covariance matrix; The computational unit calculates the algebraic rank of the class prototype covariance matrix; the number of attention heads is determined as the larger of the sum of the algebraic ranks of the class prototype covariance matrices. When the category prototype covariance matrix has non-zero feature directions, the calculation unit reads the corresponding unit column feature vectors in descending order of non-zero feature values; When equal non-zero eigenvalues ​​exist, the computation unit reads the standard basis vectors sequentially according to the coordinate dimensions of the corresponding feature subspace and projects the standard basis vectors onto the feature subspace. If the projection result is a zero vector, the standard basis vector is skipped. If the projection result is a non-zero vector, orthogonalization is performed and the direction of the eigenvector is fixed so that its first non-zero component is positive. For each non-zero feature direction, the computational unit multiplies the corresponding unit column feature vector with its transpose row vector to generate the projection matrix of the corresponding attention head; When the category prototype covariance matrix degenerates to zero rank, the computation unit divides the identity matrix corresponding to the final vector dimension by the final vector dimension, and uses it as the dimension-normalized unit projection matrix of the unique attention head. For each news text, each category, and each attention head, the computing unit reads the RNIN hidden state at each word position, the projection matrix corresponding to the attention head, and the order space category prototype of the news text corresponding to the category. The computational unit right-multiplies the RNIN hidden state by the projection matrix, and then performs an inner product with the transpose vector of the order space category prototype to obtain the category-related score. Under the same news text, the same category, and the same attention head, the calculation unit reads the category-related scores of all word positions and takes the maximum value as the central benchmark; The calculation unit subtracts the centralized benchmark from the category-related score of each position and then performs a natural index calculation to obtain the centralized index value; it sums all the centralized index values ​​under the same news text, the same category, and the same attention head to obtain the attention normalization denominator. The computational unit divides the centralization index value of each position by the attention normalization denominator to obtain the attention weight of that position; it multiplies the RNIN hidden state of each position by the corresponding attention weight and sums them along the word positions to obtain the category prototype multi-head attention feature of the news text facing the category and under the attention head. In this step, the number of attention heads is determined by the algebraic rank of the class prototype covariance matrix; when the class prototype covariance matrix degenerates, a dimension-normalized unit projection matrix is ​​used, which is determined by the vector dimension. This step outputs the category prototype, the order space category prototype, the attention weights, and the multi-head attention features of the category prototype, which can be used for subsequent MLP classification calculations. By averaging the mixed word vectors of category name lexical units to generate category prototypes, and then forming order space category prototypes through RNIN whitening projection matrix and layer normalization, short category names can be transformed into category-side semantic objects that can be compared with the hidden state of news. The number of attention heads is then determined by the algebraic rank of the category prototype covariance matrix, and a projection matrix is ​​generated based on the non-zero feature direction, so that the difference direction between different categories participates in multi-head attention aggregation. The centralization index normalization process further reduces the dominance of large value scores on weight distribution, thereby enabling category name semantics to participate in the key position screening of news text. Step 8: Output news categories by combining MLP (Multilayer Perception Network) with geometric energy; Reference Figure 5 In this step, the category prototype, news combination feature and category prototype multi-head attention feature are expressed in the whitening feature space, and the final classification score is formed by category evidence quantity and normalized geometric energy. The category prototype is used to represent the category-side semantic center, the news combination feature is used to represent the comprehensive semantic state of the news text to be classified, and the geometric energy is used to characterize the spatial deviation between the news combination feature and the category prototype. The MLP multilayer perceptron in this step consists of news combination features, category anchoring features, news combination feature covariance matrix, whitening weight matrix and error function type nonlinear activation operation, without calling external MLP training model parameters; The computational unit performs an arithmetic mean on the local reversible interaction features of the CINN along the word positions to obtain the local interaction mean. The computational unit performs an arithmetic average of the RNIN hidden states along the word positions to obtain the sequential state mean. The computing unit sequentially concatenates the local interaction mean, the sequential state mean, the multi-head attention features of the category prototypes under each attention head, and the sequential space category prototypes to generate news combination features; The computing unit expands the sequential space category prototype in the same dimension according to the composition structure of the news combination features to form category anchoring features; The same-dimensional expansion includes: copying the sequence space category prototype once to the position corresponding to the local interaction mean, copying the sequence space category prototype once to the position corresponding to the sequence state mean, copying the sequence space category prototype once to the position corresponding to each attention head, and copying the sequence space category prototype again to the position corresponding to the sequence space category prototype, thereby making the category anchoring feature and the news combination feature have the same dimension. The calculation unit summarizes the combined features of all news texts and all categories, and calculates the mean vector of the combined news features. For each news combination feature, the computing unit calculates its deviation row vector relative to the mean vector of the news combination feature; the transpose column vector of the deviation row vector is multiplied by the deviation row vector to obtain the corresponding outer product matrix; The calculation unit sums all the outer product matrices and divides them by the product of the number of news texts and the number of categories to obtain the news combination feature covariance matrix. The computational unit adds a zero-smoothing factor to the news combination feature covariance matrix and multiplies it by the identity matrix of the same dimension to obtain a stabilization term, and performs principal value symmetric inverse square root processing to generate the whitening weight matrix of the MLP multilayer perceptron network. The calculation unit subtracts the mean vector of the news combination features from the news combination features and then multiplies it by the whitening weight matrix on the right to obtain the whitened news combination features; The calculation unit subtracts the mean vector of the news combination features from the category anchor features and then multiplies it by the whitening weight matrix to obtain the whitening category anchor features. The computing unit performs an error function-type nonlinear activation operation on each whitened news combination feature; the error function-type nonlinear activation operation includes: dividing each component of the whitened news combination feature by the square root of 2 and inputting it into an error function; adding 1 to the output value of the error function and multiplying it by one-half; multiplying the result element-by-element with the corresponding component of the whitened news combination feature to obtain the activated news combination feature; The computational unit performs an inner product between the whitening category anchor feature and the activated news combination feature to obtain the category evidence molecule; The calculation unit multiplies the L2 norm of the whitening category anchor feature with the L2 norm of the activated news combination feature, and adds it to the zero-prevention smoothing amount to obtain the denominator of the category evidence. The calculation unit divides the numerator of the category evidence by the denominator of the category evidence to obtain the category evidence quantity; For the same news text, the calculation unit reads the category evidence quantity of all its corresponding categories and calculates the average category evidence quantity; the average value is then subtracted from the category evidence quantity to obtain the category evidence deviation. The calculation unit calculates the square mean of the deviations of all categories of evidence, adds it to the zero-prevention smoothing amount, and takes the square root to obtain the standardized denominator of the category evidence. The calculation unit divides the deviation of each category of evidence by the standardized denominator of the category evidence to obtain the standardized category evidence quantity; For the same news text and the same category, the computing unit reads the category prototype multi-head attention features corresponding to all attention heads, and performs an arithmetic mean on these features to obtain the mean of the multi-head attention features; The computing unit calculates the difference vector between the mean of the multi-head attention features and the corresponding order space category prototype, and calculates the square norm of the difference vector to obtain the geometric deviation of the category prototype. For the same news text, the computing unit reads the category prototype geometric deviation of all categories, calculates the arithmetic mean of these geometric deviations, and adds it to the zero-prevention smoothing amount to obtain the geometric energy normalized denominator; The calculation unit divides the geometric deviation of the category prototype for each category by the normalized denominator of the geometric energy to obtain the normalized geometric energy. The computational unit subtracts the normalized geometric energy from the standardized category evidence quantity to obtain the final classification score; The calculation unit compares the final classification scores of the same news text for each category and outputs the category with the highest final classification score as the news category; when multiple categories have the same final classification score, the category with the smallest index is output. By performing positional mean processing on the local reversible interaction features of CINN and the hidden state of RNIN respectively, and concatenating them with the multi-head attention features of category prototypes and the category prototypes in the order space to form news combination features, the local structure, order state, category attention information and category prototype information are jointly entered into the MLP evidence compression process. The whitening weight matrix and error function type nonlinear activation make the category evidence quantity comparable in the same whitening space. Furthermore, the geometric deviation between the multi-head attention features of category prototypes and the category prototypes in the order space is combined to generate the final classification score, so that the classification output is simultaneously constrained by the evidence strength and the consistency of the category space. Example 2: A system for implementing the deep learning-based intelligent news text classification method, the system operation interface of which is as follows. Figures 6-13 As shown, it mainly includes the following modules: Input preprocessing module: used to collect news text and category names, and generate a unified corpus and vocabulary; Semantic representation module: used to count word co-occurrence relationships and generate context semantic vectors; Sub-word representation module: used to aggregate character fragment information and generate character sub-word vectors; Hybrid representation module: used to generate hybrid word vectors based on context semantic vectors and character sub-word vectors; Sequence matrix module: used to perform positional encoding on mixed word vectors and generate a second-order energy matrix; CINN Interactive Module: Used to perform reversible interactive transformations on the second-order energy matrix and output locally reversible interactive features; RNIN interaction module: used to generate hidden states based on local reversible interaction features; Attention aggregation module: used to generate category prototypes based on hidden states and aggregate multi-head attention features based on category prototypes; Classification output module: Used to generate category evidence based on multi-head attention features and output news categories.

[0030] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0031] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent classification method for news text, characterized in that, include: Establish a text input terminal, a category input terminal, a normalization unit, a segmentation unit, a character splitting unit, a calculation unit, and an output terminal to collect news text and category names; The news texts and category names are subjected to unified encoding, character standardization, word segmentation, and character fragment splitting to form a unified corpus and vocabulary; Count the number of lexical terms and the number of undirected neighbor co-occurrences, construct the context co-occurrence intensity matrix and the stable semantic matrix, and generate the Word2vec context semantic vector; Based on the character fragments and word frequency, the Word2vec context semantic vectors of the same character fragments are weighted and aggregated to generate FastText character sub-word vectors; A hybrid gating quantity is generated based on word frequency scarcity and character fragment novelty, and then the Word2vec context semantic vector and FastText character sub-word vector are fused to obtain a hybrid word vector; By superimposing positional basis functions on the hybrid word vectors, a hybrid representation matrix of the news text is formed. The dimensions are then split, and local contextual mean and second-order energy matrix are constructed. Based on the second-order energy matrix, scaling and translation branches of the CINN locally invertible interaction network are generated to obtain locally invertible interaction features, which are then used to generate hidden states through the RNIN sequential normalization interaction network. Category prototypes are generated from the hybrid word vectors of category name lexics. Multi-head attention features of the category prototypes are constructed, input into an MLP (Multilayer Perceptron) network, and combined with geometric energy to output news categories.

2. The intelligent news text classification method based on deep learning according to claim 1, characterized in that, The process of establishing a text input terminal, a category input terminal, a normalization unit, a segmentation unit, a character splitting unit, a calculation unit, and an output terminal to collect news text and category names specifically includes: Receive the news text to be categorized through the text input terminal, and receive the preset category name through the category input terminal; The news text and category names are standardized by performing unified encoding, case normalization, full-width and half-width character normalization, and punctuation boundary preservation. The news text word sequence and category name word set are generated separately by segmentation units; Each word is used to generate a set of continuous character segments by a character splitting unit, and the set of continuous character segments contains at least the corresponding word itself. The news text and category names are merged into a unified corpus by the computing unit, and a unified vocabulary is formed based on the word units that appear in the unified corpus. The news categories output by the computing unit are received at the output terminal.

3. The intelligent news text classification method based on deep learning according to claim 1, characterized in that, The steps of calculating the number of lexical terms and the number of undirected adjacent co-occurrences, constructing a context co-occurrence strength matrix and a stable semantic matrix, and generating a Word2vec context semantic vector specifically include: Traverse the sequence of word elements in the news text and the set of word elements for category names in the unified corpus, and count the number of occurrences of each word element; Count the number of undirected co-occurrences of any two words based on their first-order adjacency relationship; The frequency of word occurrences, the frequency of word co-occurrences, the total number of co-occurrences, and the size of the vocabulary are smoothed and normalized to generate a context co-occurrence intensity matrix. A stable semantic matrix is ​​obtained by adding a stabilization term determined by the total number of co-occurrences and the vocabulary size to the context co-occurrence intensity matrix. Perform singular value decomposition on the stable semantic matrix and determine the retained dimension based on the direction of the singular values ​​that is not lower than the mean of the singular values; The retained directions are sorted in descending order of singular values. When there are equal singular values, the standard basis vectors are projected into a subspace according to the unified vocabulary index order, skipping zero projection. Non-zero projections are orthogonalized and the vector directions are fixed. Extract the Word2vec context semantic vector for each word.

4. The intelligent news text classification method based on deep learning according to claim 1, characterized in that, The step of weighted aggregation of Word2vec context semantic vectors of the same character segment based on character fragments and word frequency to generate FastText character sub-word vectors specifically includes: Read the set of consecutive character segments generated by the character splitting unit; Retrieve lexicons containing the same character segment from a unified thesaurus; Treat the lexical units containing the same character segment as aggregate objects; Using the frequency of occurrence of each aggregated object in the unified corpus as the weighting criterion, a weighted average is performed on the Word2vec context semantic vector of each aggregated object to generate the FastText character fragment vector of the corresponding character fragment; The FastText character fragment vectors contained in the target word are averaged to generate FastText character sub-word vectors for the target word. The FastText character sub-word vectors are kept to have the same dimension as the Word2vec context semantic vectors of the target words.

5. The intelligent news text classification method based on deep learning according to claim 1, characterized in that, The step of generating a hybrid gating quantity based on word frequency scarcity and character fragment novelty, and fusing Word2vec context semantic vectors and FastText character sub-word vectors to obtain a hybrid word vector specifically includes: Extract the highest word frequency and the median word frequency based on the occurrence frequency of each word element in the unified vocabulary; The character segments contained in words that appear at least once the median frequency are merged into a common character segment set; Term frequency scarcity is generated based on the ratio between the number of occurrences of the target term and the highest term frequency. The novelty of a character fragment is generated based on the overlap ratio between the target word metacharacter fragment set and the common character fragment set. The gating base quantity is obtained by summing the word frequency scarcity and the character fragment novelty. Using the basic gating quantity as the numerator of the occupancy ratio and the sum of the basic gating quantity and the fixed dimensionless normalized benchmark quantity as the denominator of the occupancy ratio, the occupancy ratio of FastText character sub-word vectors is extracted to obtain the hybrid gating quantity. The proportion of FastText character sub-word vectors is controlled by a hybrid gating variable, and the remaining proportion of Word2vec context semantic vectors is used for complementary fusion to generate hybrid word vectors of target words.

6. The intelligent news text classification method based on deep learning according to claim 1, characterized in that, The process of superimposing positional basis functions on the hybrid word vectors to form a hybrid representation matrix of the news text, splitting dimensions, and constructing local contextual mean and second-order energy matrices specifically includes: Read the mixed word vector of each word element according to the word element sequence of the news text; Generate a sinusoidal positional basis function based on word position, news text length, and vector dimension; Calculate the position amplitude matching quantity based on the overall energy of the mixed word vectors and the overall energy of the position basis functions within the current news text. The positional basis functions after positional amplitude matching are superimposed on the hybrid word vectors of the corresponding word units to generate a hybrid representation matrix of news text; The news text mixed representation matrix is ​​split into a first subspace matrix and a second subspace matrix according to the dimensional order; Based on the set of first-order adjacent positions of each word, a local average is performed on the first subspace matrix to generate a local context average. Perform transpose product and length normalization on the mixed representation matrix of news text to generate a second-order energy matrix.

7. The intelligent news text classification method based on deep learning according to claim 1, characterized in that, The step of generating scaling and translation branches of the CINN locally invertible interaction network based on the second-order energy matrix to obtain locally invertible interaction features, and then generating hidden states through the RNIN sequential normalized interaction network, specifically includes: From the second-order energy matrix, extract the autocorrelation block of the first subspace and the cross-correlation block from the first subspace to the second subspace; The scaling branch matrix of the CINN locally invertible interactive network is generated based on the autocorrelation blocks of the first subspace, the anti-zero smoothing amount determined by the total co-occurrence and vocabulary size, and the cross-correlation blocks. Based on the principal value symmetric inverse square root and cross-correlation blocks of the first subspace autocorrelation block, the translation branch matrix of the CINN locally invertible interactive network is generated. Affine coupling is performed on the second subspace features by local context mean, scaling branch matrix and translation branch matrix, and then concatenated with the first subspace features to obtain CINN locally invertible interactive features; Based on the locally reversible interaction features of CINN, construct the second-order energy matrix and whitening projection matrix of the interaction features; Based on the cosine consistency of the local reversible interaction features of adjacent CINNs, the sequential memory gating quantity of RNIN is generated. The current CINN local reversible interaction features, the previous hidden state, and the RNIN sequential memory gating input are processed by layer normalization operations determined by the mean and variance to generate the RNIN hidden state at the current position.

8. The intelligent news text classification method based on deep learning according to claim 1, characterized in that, The process of generating category prototypes from hybrid word vectors of category name lexics and constructing multi-head attention features for category prototypes specifically includes: Each category name is segmented using the same word segmentation method as the news text; Read the mixed word vectors of the terms contained in the category name, and perform averaging on the mixed word vectors to generate the category prototype; Read the whitening projection matrix generated by the RNIN sequential normalized interactive network; The category prototypes are input into the whitening projection matrix and the layer normalization operation to generate the order space category prototypes. Construct a category prototype covariance matrix based on the order space category prototypes corresponding to all news texts and all categories; The number of attention heads is determined based on the algebraic rank of the category prototype covariance matrix; The projection matrix of the corresponding attention head is generated based on the non-zero feature direction. When the category prototype covariance matrix degenerates, the dimension-normalized unit projection matrix is ​​used. Calculate the category-related score based on the RNIN hidden state, projection matrix, and order space category prototype; Centralized index normalization is applied to category-related scores to obtain attention weights; The RNIN hidden states are weighted and aggregated according to the attention weights to obtain the multi-head attention features of the category prototype.

9. The intelligent news text classification method based on deep learning according to claim 1, characterized in that, The step of inputting the news category into the MLP (Multilayer Perceptron) and combining it with geometric energy specifically includes: The local reversible interaction features of CINN are averaged along the word positions to obtain the local interaction mean. The RNIN hidden states are averaged along the word positions to obtain the sequential state mean. The local interaction mean, the sequential state mean, the category prototype multi-head attention feature, and the sequential space category prototype are concatenated into a news combination feature; The sequential space category prototype is extended in the same dimension according to the composition structure of news combination features to form category anchoring features; Generate a mean vector, covariance matrix, and whitening weight matrix based on the combined features of all news items; News combination features and category anchoring features are mapped to the same whitening space, and category evidence is generated by error function-type nonlinear activation operation of MLP multilayer perceptron. Standardize the amount of evidence for each category within the same news text; Normalized geometric energy is generated based on the squared norm distance between the multi-head attention features of the category prototype and the category prototype in the order space. The difference between the standardized category evidence quantity and the normalized geometric energy is calculated to obtain the final classification score; Output the category with the highest final classification score as the news category. When multiple categories have the same final classification score, output the category with the smallest index.

10. A system employing the deep learning-based intelligent classification method for news text as described in claim 1, characterized in that, include: Input preprocessing module: used to collect news text and category names, and generate a unified corpus and vocabulary; Semantic representation module: used to count word co-occurrence relationships and generate context semantic vectors; Sub-word representation module: used to aggregate character fragment information and generate character sub-word vectors; Hybrid representation module: used to generate hybrid word vectors based on context semantic vectors and character sub-word vectors; Sequence matrix module: used to perform positional encoding on mixed word vectors and generate a second-order energy matrix; CINN Interactive Module: Used to perform reversible interactive transformations on the second-order energy matrix and output locally reversible interactive features; RNIN interaction module: used to generate hidden states based on local reversible interaction features; Attention aggregation module: used to generate category prototypes based on hidden states and aggregate multi-head attention features based on category prototypes; Classification output module: Used to generate category evidence based on multi-head attention features and output news categories.

Images