Method and device for realizing hierarchical multi-granularity label identification processing based on adaptive dynamic path pruning, processor and storage medium thereof

Abstract

The invention relates to a method for realizing hierarchical multi-granularity label identification processing based on adaptive dynamic path pruning. The method comprises the following steps: acquiring content coding features; obtaining relation coding features; updating content coding features; constructing a semantic attention matrix; and hierarchical multi-label classification is carried out. The invention also relates to a device for realizing hierarchical multi-granularity label identification processing based on adaptive dynamic path pruning, a processor and a readable storage medium thereof. By the adoption of the method and device for achieving hierarchical multi-granularity label recognition processing based on self-adaptive dynamic path pruning, the processor and the computer readable storage medium of the processor, accurate classification of hierarchical multi-label texts can be achieved according to the dynamic perception ability of a large model and the semantic representation ability of a depth map neural network. According to the model, hierarchical multi-granularity identification is optimized through a dynamic path pruning technology of a large model, and the stability of the graph neural network model is improved; the classification accuracy is improved, the robustness of data disturbance is enhanced, and the method has great innovativeness.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, in particular to the field of multi-label text classification, and specifically refers to a method, device, processor and computer-readable storage medium thereof for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning. Background Art

[0002] Hierarchical text classification (HTC) is a special multi-label text classification problem, in which the classification results are assigned to one or more nodes in the label hierarchy. Existing HTC methods mainly include flat methods, local methods and global methods. Flat methods ignore the dependencies in the label hierarchy, mainly use binary classification or deep learning methods, and regard a group of labels as the same label. Unlike flat methods, local methods focus on each label, such as training classifiers for each class, training classifiers for each parent class, training classifiers for each level, and training classifiers for each sub-level. This method avoids the situation where too much label information is lost. However, building multiple classifiers will result in too many model parameters and cause error propagation. The global method mainly obtains category information and label structure dependency information through a hierarchical perception method, that is, it focuses on the leaf label information at the bottom level, and also pays attention to other layer labels outside the leaf label, effectively integrating fine-grained label and coarse-grained label information. Flat methods and local methods are prone to cause category targets to be labeled at different levels. The accuracy of global method annotation is highly dependent on the quality of samples. For articles that lack key category information or have a lot of noise, category targets will be annotated at different levels, resulting in incorrect annotation results. In order to further improve the accuracy and robustness of annotation, and to reduce the high requirements for annotation data caused by article quality and domain background knowledge, making full use of samples with different levels of granularity labels and designing a hierarchical multi-granularity recognition algorithm that models the hierarchical semantic structure of the target is very important for improving the robustness of deep neural networks.

[0003] Graph Convolutional Network (GCN) can incorporate hierarchical dependencies into the model and effectively use the dependency structure information between label levels to characterize features. Dependency-based models can capture the implicit non-local hierarchical relationships in articles, but not all information in the dependency tree is beneficial to the label classification of articles. In order to further improve the performance of the model, some corresponding pruning strategies need to be adopted to extract the dependencies in the article, better utilize the hidden information in the dependency tree, and retain the most useful dependency information.

[0004] The attention mechanism is a method that effectively focuses on key information in the model input data, which can enhance the model's ability to capture the correlation between text and labels. In the field of multi-label text classification, the text-label joint embedding is learned through label semantics and document semantics.

[0005] Based on the idea of ​​label scoring, dynamic label path clipping of large language models (LLM) is a technology to optimize model performance in hierarchical multi-label text classification tasks. The classification accuracy and efficiency of the model can be improved by dynamically adjusting the way the model processes the label path. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method, device, processor and computer-readable storage medium thereof for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning, which has high accuracy, high efficiency and a wide range of applications.

[0007] In order to achieve the above-mentioned object, the method, device, processor and computer-readable storage medium thereof for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning of the present invention are as follows:

[0008] The method for realizing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning is mainly characterized in that the method comprises the following steps:

[0009] (1) Obtaining content coding features;

[0010] (2) Obtaining relational encoding features;

[0011] (3) Update content coding features;

[0012] (4) Constructing a semantic attention matrix;

[0013] (5) Perform hierarchical multi-label classification.

[0014] Preferably, the step (1) specifically comprises the following steps:

[0015] (1.1) Generate summary content and serialize it into an initial feature matrix;

[0016] (1.2) Using graph neural networks, we associate the initial feature vectors based on the statistical label co-occurrence relationship and use convolutional neural networks to serialize them into a global content feature matrix;

[0017] (1.3) Exploring the interactions between tags and content features through graph propagation.

[0018] Preferably, the step (2) specifically comprises the following steps:

[0019] (2.1) Use a large language model to cut the category label relationship path, build a hierarchical category semantic interaction graph, and form a category semantic relationship encoding feature matrix;

[0020] (2.2) A large language model is used to dynamically adjust the pruning threshold according to the current gradient and the historical information of the gradient, and to delete or prune redundant and non-critical weights.

[0021] Preferably, the step (3) is specifically as follows:

[0022] (3.1) The hierarchical label relationship encoding features are integrated into the content feature matrix, the dimension of the relationship encoding features is assigned to the content encoding features, and the content feature matrix is ​​updated by the matrix dot product fusion method.

[0023] Preferably, the step (4) specifically comprises the following steps:

[0024] (4.1) Construct a semantic attention matrix to characterize the eigenvalues ​​of the relationship between hierarchy and content;

[0025] (4.2) The attention mechanism is used to characterize the relationship density between graph nodes and depict the dependency relationship between all nodes.

[0026] Preferably, the step (5) specifically comprises the following steps:

[0027] The features of the output calculated by the attention module are output through a layer of feedforward neural network, and the hierarchical category probability of the output features is predicted through the softmax function to obtain the hierarchical category label of the text content.

[0028] The main feature of the device for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning is that the device comprises:

[0029] a processor configured to execute computer executable instructions;

[0030] The memory stores one or more computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the method for realizing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning are implemented.

[0031] The processor for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning has the main feature that the processor is configured to execute computer-executable instructions. When the computer-executable instructions are executed by the processor, the various steps of the method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning are implemented.

[0032] The main feature of the computer-readable storage medium is that a computer program is stored thereon, and the computer program can be executed by a processor to implement the various steps of the above-mentioned method for realizing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning.

[0033] The method, device, processor and computer-readable storage medium thereof for realizing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning of the present invention can realize accurate classification of hierarchical multi-label text based on the dynamic perception ability of the large model and the semantic representation ability of the deep graph neural network. The limitations of the existing hierarchical multi-label text classification algorithm in utilizing multi-level label samples and dependency mining are solved. The model optimizes hierarchical multi-granularity recognition and improves the stability of the graph neural network model through the dynamic path pruning technology of the large model. With the help of the ability of the large model, the summary information and co-occurrence hierarchy of the domain text are extracted, and the semantic feature encoding is strengthened. In addition, the graph neural network is used for feature encoding, the text summary and dependency information are integrated, and the original information is fused through the semantically guided attention mechanism, which strengthens the hierarchical feature analysis. Compared with other advanced models, it not only improves the classification accuracy, but also enhances the robustness to data disturbances, and has great innovation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the structure of the device for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning according to the present invention.

[0035] Figure 2 The flowchart is an embodiment of the method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning of the present invention. DETAILED DESCRIPTION

[0036] In order to more clearly describe the technical content of the present invention, further description is given below in conjunction with specific embodiments.

[0037] The method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning of the present invention comprises the following steps:

[0038] (1) Obtaining content coding features;

[0039] (2) Obtaining relational encoding features;

[0040] (3) Update content coding features;

[0041] (4) Constructing a semantic attention matrix;

[0042] (5) Perform hierarchical multi-label classification.

[0043] As a preferred embodiment of the present invention, the step (1) specifically comprises the following steps:

[0044] (1.1) Generate summary content and serialize it into an initial feature matrix;

[0045] (1.2) Using graph neural networks, we associate the initial feature vectors based on the statistical label co-occurrence relationship and use convolutional neural networks to serialize them into a global content feature matrix;

[0046] (1.3) Exploring the interactions between tags and content features through graph propagation.

[0047] As a preferred embodiment of the present invention, the step (2) specifically comprises the following steps:

[0048] (2.1) Use a large language model to cut the category label relationship path, build a hierarchical category semantic interaction graph, and form a category semantic relationship encoding feature matrix;

[0049] (2.2) A large language model is used to dynamically adjust the pruning threshold according to the current gradient and the historical information of the gradient, and to delete or prune redundant and non-critical weights.

[0050] As a preferred embodiment of the present invention, the step (3) is specifically as follows:

[0051] (3.1) The hierarchical label relationship encoding features are integrated into the content feature matrix, the dimension of the relationship encoding features is assigned to the content encoding features, and the content feature matrix is ​​updated by the matrix dot product fusion method.

[0052] As a preferred embodiment of the present invention, the step (4) specifically comprises the following steps:

[0053] (4.1) Construct a semantic attention matrix to characterize the eigenvalues ​​of the relationship between hierarchy and content;

[0054] (4.2) The attention mechanism is used to characterize the relationship density between graph nodes and depict the dependency relationship between all nodes.

[0055] As a preferred embodiment of the present invention, the step (5) specifically comprises the following steps:

[0056] The features of the output calculated by the attention module are output through a layer of feedforward neural network, and the hierarchical category probability of the output features is predicted through the softmax function to obtain the hierarchical category label of the text content.

[0057] The device for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning of the present invention comprises:

[0058] a processor configured to execute computer executable instructions;

[0059] The memory stores one or more computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the method for realizing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning are implemented.

[0060] The processor of the present invention is used to implement hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning, wherein the processor is configured to execute computer executable instructions, and when the computer executable instructions are executed by the processor, the various steps of the method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning are implemented.

[0061] The computer-readable storage medium of the present invention stores a computer program thereon, and the computer program can be executed by a processor to implement the various steps of the above-mentioned method for realizing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning.

[0062] The present invention relates to the field of natural language processing technology in deep learning, and in particular to a hierarchical multi-label text classification model that integrates a large model and a deep neural network model. The model is based on the large model dynamic perception technology, assists the model in semantic feature encoding, extracts summary information and co-occurrence hierarchical structure of domain texts, and optimizes the model's hierarchical multi-granularity recognition ability of text content. At the same time, a deep graph neural network is used to integrate text summaries and dependency information, and the original information is fused through a semantically guided attention mechanism to strengthen the model's hierarchical feature analysis, thereby achieving accurate recognition of hierarchical multi-granularity information of text content.

[0063] In view of the knowledge limitations and accuracy problems existing in the prior art, the present invention proposes a hierarchical multi-label text classification model that integrates a large model and a deep neural network model. The model utilizes the dynamic perception ability of the large model and the semantic encoding technology of the deep graph neural network to extract the domain knowledge summary information and label hierarchy information in the article, and makes full use of samples with labels of different hierarchical granularity to model a hierarchical multi-granularity recognition algorithm model of the target hierarchical semantic structure, reduce the dependence on labeled data caused by sample quality and domain background knowledge, and improve the model labeling accuracy and robustness.

[0064] The present invention adopts a hierarchical multi-label text classification model that integrates a large model and a deep neural network model, and obtains serialized text data through data preprocessing tools such as PDF2TXT and HTML2TXT. The large model is used to summarize the domain knowledge in the input data, and the summary content is represented by GCN. After the global encoding module provides the global encoding of the domain knowledge article, the sequence encoding model is used to represent the overall content information of the article. The summary encoding and article encoding features are integrated, and then the sensitivity of the model to the hierarchical label features is improved through the self-attention mechanism. The model can adaptively capture information such as domain knowledge summary and hierarchical relationship, and improve the representation ability of the hierarchical label information features of the article.

[0065] The steps included in the method of the present invention are specifically as follows:

[0066] Step 1: Obtain content encoding features. First, use summary prompt to generate summary content and serialize it into an initial feature matrix. Then, use GCN to associate the initial feature vectors based on statistical label co-occurrence relationships. At the same time, consider the global features of the input content and use convolutional neural networks to serialize it into a global content feature matrix. Then, propagate messages through the graph to explore the interaction between labels and content features.

[0067] Step 2: Obtain relational encoding features. Considering the semantic interactions between hierarchical category labels, LLM is used to trim the category label relationship path, build a hierarchical category semantic interaction graph, and form a category semantic relationship encoding feature matrix. The initial input is the hierarchical label relationship matrix, and then LLM is used to dynamically adjust the trimming threshold according to the current gradient and the historical information of the gradient, trying to delete or prune redundant and non-critical weights.

[0068] Step 3: Update content encoding features. Integrate the hierarchical label relationship encoding features into the content feature matrix, reshape the dimension of the relationship encoding features to be consistent with the content encoding features, and update the content feature matrix through the matrix dot product fusion method. Reshaping here means giving a new shape to the feature (data format).

[0069] Step 4: Construct a semantic attention matrix. After updating the content feature matrix, construct a semantic attention matrix to characterize the eigenvalues ​​of the relationship between the hierarchy and the content. The attention mechanism is used to characterize the relationship density between graph nodes and characterize the dependency relationship between all nodes, thereby improving the accuracy of node relationship category judgment.

[0070] Step 5: Hierarchical multi-label classification. The features of the output of the attention module are calculated and output through a layer of feedforward neural network. The hierarchical category probability of the output features is predicted through the softmax function. Finally, the hierarchical category label of the text content is obtained.

[0071] In the specific implementation mode of the present invention, see the attached examples. Figure 2 Taking hierarchical multi-label text classification in the field of cybercrime evidence as an example, the hierarchical multi-label text classification method of the present invention includes the following steps:

[0072] 1. LLM generates summary sequence. Use summary prompt to generate summary content X′ and sequence it into feature matrix f I ∈R n ×m , where n and m are the length and number of channels of the summary content feature respectively. The formula is,

[0073] f I =f cnn (X′)……(1)

[0074] Among them, f cnn (·) is a feature extractor implemented by a convolutional neural network. For each category c, a d-dimensional semantic embedding vector is extracted using the pre-trained model GloVe.

[0075] x c =f g (w c )……(2)

[0076] Among them, w c is a category semantic word. Next, we introduce a transformation matrix W x , f ∈R d×n , the formula for integrating category semantic features is,

[0077] f c =x c W x , f f……(3)

[0078] 2. GCN content encoding. GNN is introduced to associate these vectors in the form of a graph, which is constructed based on statistical label co-occurrence. Messages are propagated through the graph to explore the interaction between labels and content features. 1 , …, x n ] represents a sentence, where x i Represents the i-th word. GCN represents this node through the category co-occurrence relationship. For example, if the words in a sentence are represented as nodes in a graph, and the graph contains n nodes, then the graph can be represented as an n×n co-occurrence matrix A. ij and A ji Represents the co-occurrence relationship between nodes i and j. In the L-layer GCN, given the input set The purpose is to output The output vector of node i in layer l It is represented by the node i and its adjacent nodes in the l-1th layer, and the formula is,

[0079]

[0080] Among them, W (l) is a weight matrix for linear transformation, b (l) is a bias vector.

[0081] After L layers of GCN process each node vector, we get the hidden representation of the node. Using these word representations, we can get the feature representation of a sentence. The formula is:

[0082]

[0083] o text =f(f (L) )=f(GCN(f (0) ))……(6)

[0084] Among them, f (L) represents all the hidden representations of the L layer, L is a hyperparameter, and the function f(·) transforms n vectors into a content feature vector containing relevant categories.

[0085] 3. Global content encoding. For the text content X, a convolutional neural network is used to serialize it into a global content feature matrix f I′ ∈R n′×m′ , where n′ and m′ are the length and number of channels of the global content feature respectively. The formula is,

[0086] f I′ =f Cnn (X)……(7)

[0087] Next, f I′ With o tect The dimensions of are reshaped to be consistent, and are merged together using the cumulative refuse method. The formula is,

[0088] o t ' ext =f I′ +o text ……(8)

[0089] 4. Hierarchical relationship encoding. For the category hierarchical relationship matrix B∈R q×q Initialization, where q represents the number of hierarchical categories, and B after reshape ij and B jiRepresents the association between categories i and j. Then, LLM dynamically adjusts the clipping threshold according to the current gradient and the historical information of the gradient, tries to delete or prune redundant and non-critical weights, and gives a score of 1 or 0 during dynamic clipping. If not clipped, it is assigned a score of 1, and if clipped, it is assigned a score of 0, and updates the weight of the feature matrix. Then, the output vector of node i in the lth layer of GCN is The formula represented by node i and its adjacent nodes in layer l-1 is updated as follows,

[0090]

[0091] 5. Semantic attention. Semantic attention matrix P∈R n×n , focusing on the dependency relationship between nodes. For example, the relationship between node i and node j is represented by P ij or P ji Represents. Initially, random numbers are used to fill the P value. Then the self-attention mechanism is used to generate the relationship feature value. Next, matrix multiplication is performed on the attention matrix and the original feature. Finally, the multiplied matrix is ​​element-wise summed with the original feature to obtain the global dependency between nodes. The closer the relationship between two nodes, the greater the impact on this value. The formula is as follows,

[0092]

[0093] Next, the attention matrix is ​​multiplied by the original node feature. Finally, the result is multiplied by a learning factor β and added element by element with the original feature to obtain the final representation f ′(L) , the formula is as follows,

[0094]

[0095] Among them, β is gradually learned from 0. The final representation obtained contains the dependency information between all nodes, thereby improving the accuracy of judging the node relationship category.

[0096] The formula of content feature vector is updated as follows:

[0097] o text =f(attention·f′ (L) )……(12)

[0098] 6. Hierarchical multi-label category discrimination. The features output by the attention module are represented by the output relational features through the function f(·), where f(·) is a layer of feedforward neural network. Then, the output features are predicted by the softmax function to predict the probability P of the hierarchical category. The formula is as follows:

[0099] P(y ij |o text )=softmax(MLP(yij ))……(13)

[0100] Among them, MLP(·) is a multi-layer perceptron. Here, the data set has 2 levels and 4 labels. The probability distribution calculation outputs {0.9, 0.064, 0.012, 0.024}, and the value corresponding to the category label of "electronic data-electronic chat" is the largest, that is, the hierarchical multi-label category of the text content output by the model is "electronic data-electronic chat."

[0101] The technical solution of the present invention does not directly use the attention mechanism to encode the global features of the text, but instead uses a graph structure feature representation method to form an adjacency matrix of key content information and label information in the form of nodes and edges. Then, the attention mechanism encoding is adopted on the basis of the adjacency matrix. This method greatly improves the key feature representation capability between text content information and multi-granularity label relationship information while reducing the semantic encoding capability of the sparse feature matrix.

[0102] The technical solution of the present invention is based on the summary and dynamic perception capabilities of the large model, and there is no strong dependence between the prompt and the model training. This technical solution can adaptively condense the key information in the text, with weak manual intervention and stronger model generalization ability.

[0103] The main steps of the technical solution of the present invention include obtaining content coding features, obtaining relationship coding features, updating content coding features, constructing a semantic attention matrix, performing hierarchical multi-label classification, etc.

[0104] The specific implementation scheme of this embodiment can refer to the relevant description in the above embodiment, which will not be repeated here.

[0105] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0106] It should be noted that, in the description of the present invention, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "plurality" refers to at least two.

[0107] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention belong.

[0108] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution device. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0109] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the corresponding program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0110] In addition, each functional unit in each embodiment of the present invention may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0111] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0112] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0113] The method, device, processor and computer-readable storage medium thereof for realizing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning of the present invention can realize accurate classification of hierarchical multi-label text based on the dynamic perception ability of the large model and the semantic representation ability of the deep graph neural network. The limitations of the existing hierarchical multi-label text classification algorithm in utilizing multi-level label samples and dependency mining are solved. The model optimizes hierarchical multi-granularity recognition and improves the stability of the graph neural network model through the dynamic path pruning technology of the large model. With the help of the ability of the large model, the summary information and co-occurrence hierarchy of the domain text are extracted, and the semantic feature encoding is strengthened. In addition, the graph neural network is used for feature encoding, the text summary and dependency information are integrated, and the original information is fused through the semantically guided attention mechanism, which strengthens the hierarchical feature analysis. Compared with other advanced models, it not only improves the classification accuracy, but also enhances the robustness to data disturbances, and has great innovation.

[0114] In this specification, the present invention has been described with reference to specific embodiments thereof. However, it is apparent that various modifications and variations may be made without departing from the spirit and scope of the present invention. Therefore, the specification and drawings should be regarded as illustrative rather than restrictive.

Claims

1. A method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning, characterized in that: The method comprises the following steps: (1) Obtaining content coding features; (2) Obtaining relational encoding features; (3) Update content coding features; (4) Constructing a semantic attention matrix; (5) Perform hierarchical multi-label classification.

2. The method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning according to claim 1 is characterized in that: The step (1) specifically comprises the following steps: (1.1) Generate summary content and serialize it into an initial feature matrix; (1.2) Using graph neural networks, we associate the initial feature vectors based on the statistical label co-occurrence relationship and use convolutional neural networks to serialize them into a global content feature matrix; (1.3) Exploring the interactions between tags and content features through graph propagation.

3. The method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning according to claim 1 is characterized in that: The step (2) specifically comprises the following steps: (2.1) Use a large language model to cut the category label relationship path, build a hierarchical category semantic interaction graph, and form a category semantic relationship encoding feature matrix; (2.2) A large language model is used to dynamically adjust the pruning threshold according to the current gradient and the historical information of the gradient, and to delete or prune redundant and non-critical weights.

4. The method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning according to claim 1, characterized in that: The step (3) is specifically as follows: (3.1) The hierarchical label relationship encoding features are integrated into the content feature matrix, the dimension of the relationship encoding features is assigned to the content encoding features, and the content feature matrix is ​​updated by the matrix dot product fusion method.

5. The method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning according to claim 1 is characterized in that: The step (4) specifically comprises the following steps: (4.1) Construct a semantic attention matrix to characterize the eigenvalues ​​of the relationship between hierarchy and content; (4.2) The attention mechanism is used to characterize the relationship density between graph nodes and depict the dependency relationship between all nodes.

6. The method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning according to claim 1, characterized in that: The step (5) specifically comprises the following steps: The features of the output calculated by the attention module are output through a layer of feedforward neural network, and the hierarchical category probability of the output features is predicted through the softmax function to obtain the hierarchical category label of the text content.

7. A device for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning, characterized in that: The device comprises: a processor configured to execute computer executable instructions; A memory storing one or more computer executable instructions, wherein when the computer executable instructions are executed by the processor, the steps of the method for realizing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning as described in any one of claims 1 to 6 are implemented.

8. A processor for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning, characterized in that: The processor is configured to execute computer executable instructions. When the computer executable instructions are executed by the processor, the steps of the method for implementing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning as described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program can be executed by a processor to implement the various steps of the method for realizing hierarchical multi-granularity label recognition processing based on adaptive dynamic path pruning as described in any one of claims 1 to 6.

Images