Deep Learning-Based Text Key Feature Extraction System and Method

Abstract

This invention relates to the field of text extraction, specifically to a deep learning-based system and method for extracting key text features. The system includes a text filtering module, a graph construction module, a deep learning module, a cluster feature module, and an input fusion module. The text filtering module merges text clusters, the graph construction module generates a stereo model, the deep learning module expands the model using clustered text, the cluster feature module prunes the stereo model, and the input fusion module captures new text features to generate fused text. This invention provides a visual interface for training complex text, enhances the machine's ability to analyze the coherence of long texts, improves the generalization ability of machine learning, is suitable for processing large amounts of complex, high-dimensional text data, reduces external data requirements, adapts to changes in data distribution, reduces the computational burden on the model, and achieves stable language frameworks and precise fusion of text features.

Description

Technical Field

[0001] This invention relates to the field of text extraction, specifically to a text key feature extraction system and method based on deep learning. Background Technology

[0002] Text feature extraction refers to the process of identifying and extracting the text features that best represent the core theme from unstructured raw text data. Through text feature extraction, human natural language can be transformed into a machine-readable simplified digital form while preserving the core information of the text to the greatest extent possible. Feature extraction is one of the fundamental steps in natural language processing and text processing, and is usually accomplished by machine learning algorithms through iterative dimensionality reduction and simplification of the text.

[0003] In machine learning tasks, using the full-text vocabulary for machine training leads to poor model performance and long training times, while selecting high-frequency words for training can easily result in the loss of textual information. Feature extraction can shorten the machine learning process, but most text features are complex, with the relationships between fields and words forming a vast vector network. Existing feature extraction methods are based on short texts, and machine learning's text processing capabilities struggle to handle large clusters of text.

[0004] In addition, commonly used text processing models are based on databases for text matching and text operations. However, in different environments and situations, when there is enough information in the text cluster, the vocabulary in the database cannot be well adapted to the user's text, resulting in problems such as low information acquisition efficiency and poor model training matching. Summary of the Invention

[0005] The purpose of this invention is to provide a text key feature extraction system and method based on deep learning to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a text key feature extraction system based on deep learning, comprising: a text filtering module, a graph construction module, a deep learning module, a cluster feature module, and an input fusion module;

[0007] The text filtering module is used to cluster, deduplicatize, and merge texts from the same information source using topic models or semantic similarity models to obtain text clusters. It calculates the average semantic similarity between each text in the cluster and all other texts, selects the text with the highest average as the feature text, and uses part-of-speech tagging and text recognition technology to perform NLP analysis on the feature text to extract fields from the feature text, including: noun phrases, verb phrases, and named entities.

[0008] The graph construction module is used to count the TF-IDF, TextRank, or BERT values ​​of each field, select core fields with values ​​higher than the threshold as initial 3D graph nodes, count the co-occurrence relationship between every two core fields in the feature text, calculate the frequency of field co-occurrence or point mutual information value, obtain the edge weights between two nodes after scaling, establish the field model of the feature text, input the field model into deep learning software, use graph neural networks or graph embedding algorithms to perform free fractals in 3D space, map the field nodes and edge weights to a high-dimensional vector space, and form a 3D graph model that reflects the semantic relationship between fields.

[0009] The deep learning module is used to traverse other texts within the text cluster, extract core fields from other texts, and if the core field already exists in the model, then the text is connected to the corresponding node of the existing model. If the core field is not in the model, but the TF-IDF, TextRank, or BERT value of the field in the text cluster exceeds a set threshold, then it is added to the model as a new node, and the node edge weights are assigned, and connection edges are established with existing nodes, until the text cluster forms a stable three-dimensional space.

[0010] The cluster feature module is used to learn the extended node model through graph convolutional network, determine the stable values ​​of model nodes and edges, identify and remove weak nodes with strength lower than the stable value and weak edges with edge weight lower than the stable value in the model, form a stable feature framework for the current text cluster, use a pre-trained model to generate a model, take the node with the highest centrality in the feature framework and the corresponding field as conditional input, output key feature text and transmit it to the user.

[0011] The input fusion module is used to convert new text from an unknown source into a stereo graph model structure when it receives new text, calculate its graph similarity or semantic similarity with the feature frames of each cluster, and if the similarity is higher than the threshold, it is considered that the new text has been captured by the current cluster's frame. Sub-model features are extracted according to the captured structure, and all sub-model features are input into the generative model to generate fused text.

[0012] Furthermore, the text filtering module includes: an information classification unit and a field extraction unit;

[0013] The information classification unit is used to process the original text corpus in the database and divide the text clusters according to the information source;

[0014] The field extraction unit is used to remove irrelevant characters, stop words, and standardized formats from the text, extract feature fields, and represent them as semantic vectors.

[0015] Furthermore, the graph construction module includes: a vector clustering unit, a node simulation unit, and a graph model unit;

[0016] The vector clustering unit is used to cluster semantic vectors using a clustering algorithm;

[0017] The node simulation unit is used to establish the core fields extracted from the feature text as graph nodes and initialize the edge weights of the nodes.

[0018] The graph model unit is used to use edge weights as the attraction strength between nodes, allowing nodes to be freely arranged in three-dimensional space to form a three-dimensional graph model.

[0019] Furthermore, the deep learning module includes: a feature traversal unit and a model expansion unit;

[0020] The feature traversal unit is used to extract the core fields of other texts in the cluster, add them as new nodes to the graph, and assign edge weights to the existing nodes.

[0021] The model expansion unit is used to connect the field nodes of other texts in the cluster to the current model one by one until the model is stable.

[0022] Furthermore, the cluster feature module includes: a frame pruning unit and a text generation unit;

[0023] The frame pruning unit is used to remove redundant or noisy nodes with low weight and weak connections to obtain a stable feature frame.

[0024] The text generation unit is used to generate standard text output based on the feature framework. The output text is used for machine digitization learning and text cluster generalization.

[0025] Furthermore, the input fusion module includes: a model capture unit, a text segmentation unit, and a feature fusion unit;

[0026] When the new text is captured by a single frame, the model capture unit extracts independent features from the new text and feeds the independent features as candidate nodes back to the feature model.

[0027] The text splitting unit is used to split the new text according to the node connection state when the new text is captured by more than one frame, so that each sub-model has the highest matching degree with the feature frame that captured it.

[0028] The feature fusion unit is used to recombine the processed new text features into a fused text output based on the text generation model and the logical order of each text cluster.

[0029] A deep learning-based method for extracting key text features includes the following steps:

[0030] Step S1. Merge texts from the same information source into a text cluster. Use the text with the highest average semantic similarity within the text cluster as the feature text. Perform NLP analysis on the feature text to extract all fields from the feature text.

[0031] Step S2. Select core fields in the feature text whose clustering index is higher than the threshold as initial graph nodes, calculate the co-occurrence frequency or point mutual information value between every two core fields, obtain the edge weight between the two nodes after scaling, and map the field nodes and edge weights to a high-dimensional vector space to form a 3D graph model.

[0032] Step S3. Extract the core fields of other texts in the cluster, add them as new nodes to the graph, and establish edge weights with existing nodes. Connect the field nodes of other texts in the cluster to the current model one by one until the model forms a stable three-dimensional space.

[0033] Step S4. Learn the extended node model through the graph convolutional network, determine the stable values ​​of the model nodes and edges, identify and remove weak nodes with strength lower than the stable value and weak edges with edge weights lower than the stable value in the model, obtain a stable feature frame, and generate standard text based on the feature frame;

[0034] Step S5. When a new text is received, the new text is converted into a stereo model structure, and its similarity with the feature frames of each cluster is calculated. If the similarity is higher than the threshold, it means that the new text has been captured by the current cluster's frame. Sub-model features are extracted according to the captured structure, fused text is generated, and the feature frames of each cluster are updated.

[0035] Furthermore, step S1 includes:

[0036] Step S11. Process the original text corpus in the database, remove irrelevant characters, stop words and standardized formats from the corpus text, and use topic model or semantic similarity model to cluster, deduplicatize and merge texts from the same information source to obtain text clusters;

[0037] Step S12. Calculate the average semantic similarity between each text in the cluster and all other texts, select the text with the highest average as the feature text, and use part-of-speech tagging and text recognition technology to perform NLP analysis on the feature text to extract fields from the feature text, including: noun phrases, verb phrases and named entities.

[0038] Furthermore, step S2 includes:

[0039] Step S21. Cluster the semantic vectors in the feature text using a clustering algorithm. Calculate the TF-IDF, TextRank, or BERT value of each field according to the clustering results. Select core fields with values ​​higher than the threshold as initial 3D graph nodes. Statistically analyze the co-occurrence relationship between every two core fields in the feature text. Calculate the frequency of field co-occurrence or point mutual information value. After scaling, obtain the edge weight between the two nodes.

[0040] Step S22. Input the initial 3D graphic nodes and the edge weights of each node into the deep learning software, and use graph neural networks or graph embedding algorithms to freely fractalize in 3D space. Use the edge weights as the attraction strength between nodes, and let the nodes be freely arranged in 3D space to form a 3D graph model that reflects the semantic relationship between fields.

[0041] Furthermore, step S3 includes:

[0042] Step S31. Traverse the text within the cluster and extract the core fields from the other texts. If the core field already exists in the model, then connect the text with the corresponding node in the existing model. If the core field is not in the model, but the TF-IDF, TextRank, or BERT value of the field in the text cluster exceeds the set threshold, then add it as a new node to the model and assign edge weights to the node.

[0043] Step S32. Connect the field nodes of other texts in the cluster to the current model one by one until the text cluster forms a stable three-dimensional space.

[0044] Furthermore, step S4 includes:

[0045] Step S41. Analyze the expanded graph, and after pruning and optimization, remove redundant or noisy nodes with low weight and weak connections, and retain the stable feature framework with core nodes and strong connections.

[0046] Step S42. Using a pre-trained generative model, the node with the highest centrality within the feature frame and its corresponding field are used as conditional inputs. The key feature text is output and transmitted to the user. The output text is used for machine digitization learning and text cluster generalization.

[0047] Furthermore, step S5 includes:

[0048] Step S51. Input new text from an unknown source, calculate the graph similarity or semantic similarity between the new text vector and the center vector of each text cluster. If the similarity is higher than the threshold, it is considered that the new text has been captured by the frame of the current cluster. Select all feature frames with similarity exceeding the threshold. If the new text is not captured, store the new text in the corpus database.

[0049] Step S52. If the new text is captured by a single frame, extract the independent features in the new text and feed the independent features as candidate nodes back to the feature model. If the new text is captured by more than one frame, split the new text according to the node connection state so that each sub-model has the highest matching degree with the feature frame that captured it, obtain the new text fragments in each feature frame, and combine them into a fused text output according to the logical order of each text cluster.

[0050] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0051] This invention merges texts from the same information source into text clusters, selects the field with the most repetition frequency in the feature text as a 3D graphic node, and uses the co-occurrence relationship between fields as the node attraction strength to establish a field model of the feature text. It provides a visual interface for training complex texts, is suitable for processing large amounts of complex and high-dimensional text data, improves the machine's ability to analyze the coherence of long texts, and enhances the generalization of machine learning.

[0052] This invention improves the assignment of node strength to text node fields, enabling text clusters to form a stable spatial model, determining the feature framework of the text cluster, and generating standard feature text output according to the feature framework. It can automatically capture a large number of text lexical, syntactic, and semantic features, adapt to text analysis tasks in specific environments, reduce external data requirements, adapt to changes in data distribution, and reduce the computational burden of the model.

[0053] This invention can capture the feature framework of each text cluster to form a new text model. After capture, the feature framework is retained to perform field deletion and new feature extraction on the new text, thereby improving the automation of feature extraction, enhancing the pruning and strengthening functions of the deep learning model, stabilizing the language framework, and achieving accurate fusion of text features through graph algorithms and generative models. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a schematic diagram of the structure of the deep learning-based text key feature extraction system of the present invention;

[0056] Figure 2 This is a schematic diagram illustrating the steps of the deep learning-based text key feature extraction method of the present invention. Detailed Implementation

[0057] 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.

[0058] Please see Figure 1 The present invention provides a technical solution: a text key feature extraction system based on deep learning, comprising: a text filtering module, a graph construction module, a deep learning module, a cluster feature module, and an input fusion module;

[0059] The text filtering module is used to cluster, deduplicatize, and merge texts from the same information source using topic models or semantic similarity models to obtain text clusters. It calculates the average semantic similarity between each text in the cluster and all other texts, selects the text with the highest average as the feature text, and uses part-of-speech tagging and text recognition technology to perform NLP analysis on the feature text to extract fields from the feature text, including: noun phrases, verb phrases, and named entities.

[0060] The text filtering module includes: an information classification unit and a field extraction unit;

[0061] The information classification unit is used to process the original text corpus in the database and divide the text clusters according to the information source;

[0062] The field extraction unit is used to remove irrelevant characters, stop words, and standardized formats from the text, extract feature fields, and represent them as semantic vectors.

[0063] The graph construction module is used to count the TF-IDF, TextRank, or BERT values ​​of each field, select core fields with values ​​higher than the threshold as initial 3D graph nodes, count the co-occurrence relationship between every two core fields in the feature text, calculate the frequency of field co-occurrence or point mutual information value, obtain the edge weights between two nodes after scaling, establish the field model of the feature text, input the field model into deep learning software, use graph neural networks or graph embedding algorithms to perform free fractals in 3D space, map the field nodes and edge weights to a high-dimensional vector space, and form a 3D graph model that reflects the semantic relationship between fields.

[0064] The graph construction module includes: a vector clustering unit, a node simulation unit, and a graph model unit;

[0065] The vector clustering unit is used to cluster semantic vectors using a clustering algorithm;

[0066] The node simulation unit is used to establish the core fields extracted from the feature text as graph nodes and initialize the edge weights of the nodes.

[0067] The graph model unit is used to use edge weights as the attraction strength between nodes, allowing nodes to be freely arranged in three-dimensional space to form a three-dimensional graph model.

[0068] The deep learning module is used to traverse other texts within the text cluster, extract core fields from other texts, and if the core field already exists in the model, then the text is connected to the corresponding node of the existing model. If the core field is not in the model, but the TF-IDF, TextRank, or BERT value of the field in the text cluster exceeds a set threshold, then it is added to the model as a new node, and the node edge weights are assigned, and connection edges are established with existing nodes, until the text cluster forms a stable three-dimensional space.

[0069] The deep learning module includes: a feature traversal unit and a model expansion unit;

[0070] The feature traversal unit is used to extract the core fields of other texts in the cluster, add them as new nodes to the graph, and assign edge weights to the existing nodes.

[0071] The model expansion unit is used to connect the field nodes of other texts in the cluster to the current model one by one until the model is stable.

[0072] The cluster feature module is used to learn the extended node model through graph convolutional network, determine the stable values ​​of model nodes and edges, identify and remove weak nodes with strength lower than the stable value and weak edges with edge weight lower than the stable value in the model, form a stable feature framework for the current text cluster, use a pre-trained model to generate a model, take the node with the highest centrality in the feature framework and the corresponding field as conditional input, output key feature text and transmit it to the user.

[0073] The cluster feature module includes: a frame trimming unit and a text generation unit;

[0074] The frame pruning unit is used to remove redundant or noisy nodes with low weight and weak connections to obtain a stable feature frame.

[0075] The text generation unit is used to generate standard text output based on the feature framework. The output text is used for machine digitization learning and text cluster generalization.

[0076] The input fusion module is used to convert new text from an unknown source into a stereo graph model structure when it receives new text, calculate its graph similarity or semantic similarity with the feature frames of each cluster, and if the similarity is higher than the threshold, it is considered that the new text has been captured by the current cluster's frame. Sub-model features are extracted according to the captured structure, and all sub-model features are input into the generative model to generate fused text.

[0077] The input fusion module includes: a model capture unit, a text segmentation unit, and a feature fusion unit;

[0078] When the new text is captured by a single frame, the model capture unit extracts independent features from the new text and feeds the independent features as candidate nodes back to the feature model.

[0079] The text splitting unit is used to split the new text according to the node connection state when the new text is captured by more than one frame, so that each sub-model has the highest matching degree with the feature frame that captured it.

[0080] The feature fusion unit is used to recombine the processed new text features into a fused text output based on the text generation model and the logical order of each text cluster.

[0081] like Figure 2 As shown, the deep learning-based text key feature extraction method includes the following steps:

[0082] Step S1. Merge texts from the same information source into a text cluster. Use the text with the highest average semantic similarity within the text cluster as the feature text. Perform NLP analysis on the feature text to extract all fields from the feature text.

[0083] Step S1 includes:

[0084] Step S11. Process the original text corpus in the database, remove irrelevant characters, stop words and standardized formats from the corpus text, and use topic model or semantic similarity model to cluster, deduplicatize and merge texts from the same information source to obtain text clusters;

[0085] Step S12. Calculate the average semantic similarity between each text in the cluster and all other texts, select the text with the highest average as the feature text, and use part-of-speech tagging and text recognition technology to perform NLP analysis on the feature text to extract fields from the feature text, including: noun phrases, verb phrases and named entities.

[0086] Step S2. Select core fields in the feature text whose clustering index is higher than the threshold as initial graph nodes, calculate the co-occurrence frequency or point mutual information value between every two core fields, obtain the edge weight between the two nodes after scaling, and map the field nodes and edge weights to a high-dimensional vector space to form a 3D graph model.

[0087] Step S2 includes:

[0088] Step S21. Cluster the semantic vectors in the feature text using a clustering algorithm. Calculate the TF-IDF, TextRank, or BERT value of each field according to the clustering results. Select core fields with values ​​higher than the threshold as initial 3D graph nodes. Statistically analyze the co-occurrence relationship between every two core fields in the feature text. Calculate the frequency of field co-occurrence or point mutual information value. After scaling, obtain the edge weight between the two nodes.

[0089] Step S22. Input the initial 3D graphic nodes and the edge weights of each node into the deep learning software, and use graph neural networks or graph embedding algorithms to freely fractalize in 3D space. Use the edge weights as the attraction strength between nodes, and let the nodes be freely arranged in 3D space to form a 3D graph model that reflects the semantic relationship between fields.

[0090] Step S3. Extract the core fields of other texts in the cluster, add them as new nodes to the graph, and establish edge weights with existing nodes. Connect the field nodes of other texts in the cluster to the current model one by one until the model forms a stable three-dimensional space.

[0091] Step S3 includes:

[0092] Step S31. Traverse the text within the cluster and extract the core fields from the other texts. If the core field already exists in the model, then connect the text with the corresponding node in the existing model. If the core field is not in the model, but the TF-IDF, TextRank, or BERT value of the field in the text cluster exceeds the set threshold, then add it as a new node to the model and assign edge weights to the node.

[0093] Step S32. Connect the field nodes of other texts in the cluster to the current model one by one until the text cluster forms a stable three-dimensional space.

[0094] Step S4. Learn the extended node model through the graph convolutional network, determine the stable values ​​of the model nodes and edges, identify and remove weak nodes with strength lower than the stable value and weak edges with edge weights lower than the stable value in the model, obtain a stable feature frame, and generate standard text based on the feature frame;

[0095] Step S4 includes:

[0096] Step S41. Analyze the expanded graph, and after pruning and optimization, remove redundant or noisy nodes with low weight and weak connections, and retain the stable feature framework with core nodes and strong connections.

[0097] Step S42. Using a pre-trained generative model, the node with the highest centrality within the feature frame and its corresponding field are used as conditional inputs. The key feature text is output and transmitted to the user. The output text is used for machine digitization learning and text cluster generalization.

[0098] Step S5. When a new text is received, the new text is converted into a stereo model structure, and its similarity with the feature frames of each cluster is calculated. If the similarity is higher than the threshold, it means that the new text has been captured by the current cluster's frame. Sub-model features are extracted according to the captured structure, fused text is generated, and the feature frames of each cluster are updated.

[0099] Step S5 includes:

[0100] Step S51. Input new text from an unknown source, calculate the graph similarity or semantic similarity between the new text vector and the center vector of each text cluster. If the similarity is higher than the threshold, it is considered that the new text has been captured by the frame of the current cluster. Select all feature frames with similarity exceeding the threshold. If the new text is not captured, store the new text in the corpus database.

[0101] Step S52. If the new text is captured by a single frame, extract the independent features in the new text and feed the independent features as candidate nodes back to the feature model. If the new text is captured by more than one frame, split the new text according to the node connection state so that each sub-model has the highest matching degree with the feature frame that captured it, obtain the new text fragments in each feature frame, and combine them into a fused text output according to the logical order of each text cluster.

[0102] Example: Ten texts sent from the same website are filtered from the database. These texts are treated as a cluster. After deduplication, the semantic similarity between each text and other texts in the cluster is calculated to determine the feature texts. Noun phrases, verb phrases, and named entities are extracted from the feature texts. Fields A, B, and C in the feature texts are determined as core fields through importance calculation. The weights of the edges AB, AC, and BC due to co-occurrence relationships are 1, 2, and 3, respectively. A, B, and C are then used as nodes in the input model. The distance between nodes A, B, and C is determined according to the edge weights to form a 3D graph model.

[0103] 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.

[0104] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A text key feature extraction method based on deep learning, characterized in that, The method includes the following steps: Step S1. Merge texts from the same information source into a text cluster. Use the text with the highest average semantic similarity within the text cluster as the feature text. Perform NLP analysis on the feature text to extract all fields from the feature text. Step S2. Select core fields in the feature text whose clustering index is higher than the threshold as initial graph nodes, calculate the co-occurrence frequency or point mutual information value between every two core fields, obtain the edge weight between the two nodes after scaling, and map the field nodes and edge weights to a high-dimensional vector space to form a 3D graph model. Step S3. Extract the core fields of other texts in the cluster, add them as new nodes to the graph, and establish edge weights with existing nodes. Connect the field nodes of other texts in the cluster to the current model one by one until the model forms a stable three-dimensional space. Step S4. Learn the extended node model through the graph convolutional network, determine the stable values ​​of the model nodes and edges, identify and remove weak nodes with strength lower than the stable value and weak edges with edge weights lower than the stable value in the model, obtain a stable feature frame, and generate standard text based on the feature frame; Step S5. When a new text is received, the new text is converted into a stereo model structure, and its similarity with the feature frames of each cluster is calculated. If the similarity is higher than the threshold, it means that the new text has been captured by the current cluster's frame. Sub-model features are extracted according to the captured structure, fused text is generated, and the feature frames of each cluster are updated. Step S3 includes: Step S31. Traverse the text within the cluster and extract the core fields from the other texts. If the core field already exists in the model, then connect the text with the corresponding node in the existing model. If the core field is not in the model, but the TF-IDF, TextRank, or BERT value of the field in the text cluster exceeds the set threshold, then add it as a new node to the model and assign edge weights to the node. Step S32. Connect the field nodes of other texts in the cluster to the current model one by one until the text cluster forms a stable three-dimensional space; Step S4 includes: Step S41. Analyze the expanded graph, and after pruning and optimization, remove redundant or noisy nodes with low weight and weak connections, and retain the stable feature framework with core nodes and strong connections. Step S42. Using a pre-trained generative model, the node with the highest centrality within the feature frame and its corresponding field are used as conditional inputs. The key feature text is output and transmitted to the user. The output text is used for machine digitization learning and text cluster generalization. Step S5 includes: Step S51. Input new text from an unknown source, calculate the graph similarity or semantic similarity between the new text vector and the center vector of each text cluster. If the similarity is higher than the threshold, it is considered that the new text has been captured by the frame of the current cluster. Select all feature frames with similarity exceeding the threshold. If the new text is not captured, store the new text in the corpus database. Step S52. If the new text is captured by a single frame, extract the independent features in the new text and feed the independent features as candidate nodes back to the feature model. If the new text is captured by more than one frame, split the new text according to the node connection state so that each sub-model has the highest matching degree with the feature frame that captured it, obtain the new text fragments in each feature frame, and combine them into a fused text output according to the logical order of each text cluster. Step S2 includes: Step S21. Cluster the semantic vectors in the feature text using a clustering algorithm. Calculate the TF-IDF, TextRank, or BERT value of each field according to the clustering results. Select core fields with values ​​higher than the threshold as initial 3D graph nodes. Statistically analyze the co-occurrence relationship between every two core fields in the feature text. Calculate the frequency of field co-occurrence or point mutual information value. After scaling, obtain the edge weight between the two nodes. Step S22. Input the initial 3D graphic nodes and the edge weights of each node into the deep learning software, and use graph neural networks or graph embedding algorithms to freely fractalize in 3D space. Use the edge weights as the attraction strength between nodes, and let the nodes be freely arranged in 3D space to form a 3D graph model that reflects the semantic relationship between fields.

2. The text key feature extraction method based on deep learning according to claim 1, characterized in that: Step S1 includes: Step S11. Process the original text corpus in the database, remove irrelevant characters, stop words and standardized formats from the corpus text, and use topic model or semantic similarity model to cluster, deduplicatize and merge texts from the same information source to obtain text clusters; Step S12. Calculate the average semantic similarity between each text in the cluster and all other texts, select the text with the highest average as the feature text, and use part-of-speech tagging and text recognition technology to perform NLP analysis on the feature text to extract fields from the feature text, including: noun phrases, verb phrases and named entities.

3. A text key feature extraction system based on deep learning, characterized in that, The system includes the following modules: a text filtering module, a graph construction module, a deep learning module, a cluster feature module, and an input fusion module; The text filtering module is used to cluster, deduplicatize, and merge texts from the same information source using topic models or semantic similarity models to obtain text clusters. It calculates the average semantic similarity between each text in the cluster and all other texts, selects the text with the highest average as the feature text, and uses part-of-speech tagging and text recognition technology to perform NLP analysis on the feature text to extract fields from the feature text, including: noun phrases, verb phrases, and named entities. The graph construction module is used to count the TF-IDF, TextRank, or BERT values ​​of each field, select core fields with values ​​higher than the threshold as initial 3D graph nodes, count the co-occurrence relationship between every two core fields in the feature text, calculate the frequency of field co-occurrence or point mutual information value, obtain the edge weights between two nodes after scaling, establish the field model of the feature text, input the field model into deep learning software, use graph neural networks or graph embedding algorithms to perform free fractals in 3D space, map the field nodes and edge weights to a high-dimensional vector space, and form a 3D graph model that reflects the semantic relationship between fields. The deep learning module is used to traverse other texts within the text cluster, extract core fields from other texts, and if the core field already exists in the model, then the text is connected to the corresponding node of the existing model. If the core field is not in the model, but the TF-IDF, TextRank, or BERT value of the field in the text cluster exceeds a set threshold, then it is added to the model as a new node, and the node edge weights are assigned, and connection edges are established with existing nodes, until the text cluster forms a stable three-dimensional space. The cluster feature module is used to learn the extended node model through graph convolutional network, determine the stable values ​​of model nodes and edges, identify and remove weak nodes with strength lower than the stable value and weak edges with edge weight lower than the stable value in the model, form a stable feature framework for the current text cluster, use a pre-trained model to generate a model, take the node with the highest centrality in the feature framework and the corresponding field as conditional input, output key feature text and transmit it to the user. The input fusion module is used to convert new text from an unknown source into a stereo graph model structure when it receives new text, calculate its graph similarity or semantic similarity with the feature frames of each cluster, and if the similarity is higher than the threshold, it is considered that the new text has been captured by the current cluster's frame. Sub-model features are extracted according to the captured structure, and all sub-model features are input into the generative model to generate fused text.

4. The deep learning-based text key feature extraction system according to claim 3, characterized in that: The text filtering module includes: an information classification unit and a field extraction unit; The information classification unit is used to process the original text corpus in the database and divide the text clusters according to the information source; The field extraction unit is used to remove irrelevant characters, stop words, and standardized formats from the text, extract feature fields, and represent them as semantic vectors.

5. The deep learning-based text key feature extraction system according to claim 4, characterized in that: The graph construction module includes: a vector clustering unit, a node simulation unit, and a graph model unit; The vector clustering unit is used to cluster semantic vectors using a clustering algorithm; The node simulation unit is used to establish the core fields extracted from the feature text as graph nodes and initialize the edge weights of the nodes. The graph model unit is used to use edge weights as the attraction strength between nodes, allowing nodes to be freely arranged in three-dimensional space to form a three-dimensional graph model.

6. The deep learning-based text key feature extraction system according to claim 5, characterized in that: The deep learning module includes: a feature traversal unit and a model expansion unit; The feature traversal unit is used to extract the core fields of other texts in the cluster, add them as new nodes to the graph, and assign edge weights to the existing nodes. The model expansion unit is used to connect the field nodes of other texts in the cluster to the current model one by one until the model is stable; The cluster feature module includes: a frame trimming unit and a text generation unit; The frame pruning unit is used to remove redundant or noisy nodes with low weight and weak connections to obtain a stable feature frame. The text generation unit is used to generate standard text output based on the feature framework. The output text is used for machine digitization learning and text cluster generalization.

7. The deep learning-based text key feature extraction system according to claim 6, characterized in that: The input fusion module includes: a model capture unit, a text segmentation unit, and a feature fusion unit; When the new text is captured by a single frame, the model capture unit extracts independent features from the new text and feeds the independent features as candidate nodes back to the feature model. The text splitting unit is used to split the new text according to the node connection state when the new text is captured by more than one frame, so that each sub-model has the highest matching degree with the feature frame that captured it. The feature fusion unit is used to recombine the processed new text features into a fused text output based on the text generation model and the logical order of each text cluster.

Images