Text propaganda intention recognition method and system based on hierarchical feature map network

Abstract

The application discloses a kind of text propaganda intention recognition method and system based on hierarchical feature map network, the method includes: based on the text to be identified after pre-processing, the sequence, syntax and semantic features of text are extracted, and the corresponding text feature map is constructed according to the extracted features, and the feature map vector is obtained;Based on the feature map vector of each text feature, the isomorphism node information in graph is propagated by using the double-layer thickening operation of residual connection, and the context feature representation of node is updated, to obtain node vector;Based on node vector, the most relevant sequence, syntax and semantic information of multi-graph is aggregated by using the three-channel integrated operation based on attention, to enhance the heterogeneous node information between graphs, to obtain node enhanced vector;The node enhanced vector of three text feature maps is pooled and fused to obtain text vector, and the final classification determination is carried out through classifier, to obtain the recognition result with higher recognition precision of the propaganda intention recognition of the text to be identified.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and system for identifying textual propaganda intent based on hierarchical feature graph networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Social media is gradually replacing traditional media as the main source of information, and news texts with inflammatory or propaganda content are spreading more explosively. However, such content is often hidden within social media news texts, influencing readers' mindsets through misleading information. These news articles use repetitive information, logical fallacies, and sentence structure manipulation to persuade readers, exaggerating social issues and swaying public opinion to influence the opinions or actions of other individuals or groups to achieve specific goals, resulting in a chaotic and volatile online environment. Therefore, effectively identifying such news has become a prerequisite for creating a clean and healthy online space.

[0004] News propaganda intent recognition is a text classification task. Existing technologies mainly include three categories of text intent recognition and classification methods: article-based classification methods, sentence-based classification methods, and multimodal classification methods. Among them:

[0005] (1) Article-based classification: This method involves a coarse-grained analysis of news articles as a whole, classifying them based on their linguistic features. In most cases, propaganda labels are usually found on news media outlets, not on individual news articles; therefore, all articles published by propaganda-oriented news media are typically labeled as propaganda-oriented. However, research shows that propaganda-oriented news media also publish non-propaganda-oriented news to increase their credibility; conversely, non-propaganda-oriented news media occasionally publish propaganda-oriented news to increase traffic. Therefore, this method suffers from noise issues and cannot accurately identify news from unfamiliar sources.

[0006] (2) Sentence-based classification methods: This method performs fine-grained analysis of specific texts in news articles, classifying them based on features extracted from the news text itself and its context. Traditional machine learning methods mainly rely on manual extraction of news content features. While this method is easy to interpret, it requires manual feature design, thus necessitating the knowledge and experience of domain experts. Compared to traditional machine learning methods, deep learning models can extract deeper text features by continuously deepening the network, but handling texts with complex grammatical structures limits the expressive power of neural networks. Transformer-based pre-trained models are widely used for identifying propaganda intent. These models capture long-distance dependencies between words and contextual information by fine-tuning large corpora. However, due to the large weights in pre-training, overfitting is prone to occur when training on smaller datasets.

[0007] (3) Multimodal classification method. This method improves the performance of propaganda recognition by jointly modeling text and images. However, the current multimodal propaganda recognition model encodes different modal information independently and lacks high-order complementary information between modalities, resulting in poor final recognition accuracy.

[0008] Graph neural networks are widely used in text classification tasks because they can capture global structural information. However, there are some limitations to using existing graph neural networks for identifying inflammatory propaganda intent, mainly the following problems:

[0009] (1) Self-media data contains a lot of noisy data, such as emoticons, punctuation marks and unreadable code, which affects data analysis;

[0010] (2) Existing graph neural network representations lack diverse inter-word dependencies, lack joint learning of information enhancement between different text graphs, and are difficult to extract complex and diverse propaganda techniques for the identification of propaganda intentions;

[0011] (3) Existing graph neural networks cannot simultaneously acquire local and global information of nodes, and cannot effectively mine hidden propaganda techniques in texts or identify propaganda news with hidden nature. Summary of the Invention

[0012] To address the shortcomings of existing technologies, this invention provides a method and system for identifying textual propaganda intent based on hierarchical feature graph networks. The method constructs a hierarchical feature graph network, extracts the sequence, syntactic, and semantic features of the text, mines the co-occurrence, semantic, and syntactic relationships of words in the text, expands the update range of node information, enables information interaction between long-distance nodes and between nodes not directly connected, and uncovers hidden propaganda techniques within the text. It aggregates the most relevant sequence, syntactic, and semantic information from multiple graphs, enhances the representation of dependencies between different words, and extracts complex and diverse propaganda technique features for identifying propaganda intent, thereby improving the accuracy of propaganda intent identification.

[0013] Firstly, this disclosure provides a method for identifying textual propaganda intent based on hierarchical feature map networks.

[0014] A method for identifying textual propaganda intent based on hierarchical feature map networks, comprising:

[0015] Obtain the text to be recognized and preprocess it.

[0016] Based on the preprocessed text to be identified, sequence features, syntactic features and semantic features of the text are extracted, and corresponding text feature maps are constructed according to the extracted features to obtain feature map vectors;

[0017] Based on the feature map vector of each text feature, a two-layer coarsening operation with residual connections is used to propagate the isomorphic node information within the graph, expand the update range of node information, update the context feature representation of the node, and obtain the node vector.

[0018] Based on node vectors, an attention-based three-channel ensemble operation is used to aggregate the most relevant sequences, syntactic and semantic information of multiple graphs, enhance the heterogeneous node information between graphs, and obtain node enhancement vectors.

[0019] Based on the node enhancement vector, the node enhancement vectors of the three text feature maps are pooled and fused to obtain the text vector;

[0020] Based on text vectors, a classifier is used to make the final classification judgment, and the recognition result of the text to be recognized is obtained.

[0021] Secondly, this disclosure provides a text propaganda intent recognition system based on hierarchical feature graph networks.

[0022] A text-based propaganda intent recognition system based on hierarchical feature map networks includes:

[0023] The text acquisition and preprocessing module is used to acquire the text to be recognized and to preprocess the text to be recognized.

[0024] The feature vector construction module is used to extract sequence features, syntactic features, and semantic features of the preprocessed text to be identified, and to construct corresponding text feature maps based on the extracted features, thereby obtaining feature map vectors.

[0025] The information propagation module is used to propagate the isomorphic node information within the graph based on the feature map vector of each text feature using a two-layer coarsening operation with residual connections, thereby expanding the update range of node information, updating the context feature representation of the node, and obtaining the node vector.

[0026] The information enhancement module is used to aggregate the most relevant sequences, syntactic and semantic information of multiple graphs based on node vectors and an attention-based three-channel ensemble operation to enhance heterogeneous node information between graphs and obtain node enhancement vectors.

[0027] The text vector acquisition module is used to perform pooling and fusion of the node enhancement vectors of three text feature maps based on the node enhancement vectors to obtain the text vectors;

[0028] The classification module is used to perform the final classification judgment based on the text vector through a classifier, so as to obtain the recognition result of the text to be recognized.

[0029] Thirdly, this disclosure also provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the steps of the method described in the first aspect.

[0030] Fourthly, this disclosure also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the steps of the method described in the first aspect.

[0031] The above one or more technical solutions have the following beneficial effects:

[0032] 1. This invention provides a method and system for identifying textual propaganda intent based on hierarchical feature graph networks. It constructs a hierarchical feature graph network to extract the sequence, syntactic, and semantic features of the text, mines the co-occurrence, semantic, and grammatical relationships of words in the text, expands the update range of node information through a two-layer coarsening operation of residual connections, realizes information interaction between long-distance nodes and between nodes that are not directly connected, deeply mines the hidden propaganda techniques in the text, identifies propaganda news with hidden propaganda, and improves the accuracy of propaganda intent identification.

[0033] 2. This invention integrates the most relevant sequences, syntactic and semantic information from multiple graphs through attention-based three-channel feature integration operations, enhances the representation of word dependencies, and thereby extracts complex and diverse propaganda techniques for the purpose of identifying propaganda intent, thus improving the accuracy of propaganda intent identification. Attached Figure Description

[0034] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0035] Figure 1 This is an overall framework diagram of the text propaganda intent recognition method based on hierarchical feature map network described in the embodiments of the present invention;

[0036] Figure 2 This is a flowchart of text preprocessing in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of syntactic feature extraction in an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the semantic feature extraction process in an embodiment of the present invention;

[0039] Figure 5 This is a flowchart of the two-layer coarsening operation of residual connections in an embodiment of the present invention;

[0040] Figure 6 This is a flowchart of the attention-based three-channel feature integration operation in an embodiment of the present invention;

[0041] Figure 7 This is a flowchart illustrating the enhancement of sequence features based on the sequence enhancement channel in an embodiment of the present invention. Detailed Implementation

[0042] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0043] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0044] Example 1

[0045] This embodiment provides a method for identifying textual propaganda intent based on hierarchical feature map networks, such as... Figure 1 As shown, the specific steps include:

[0046] Step S1: Obtain the text to be recognized and preprocess it.

[0047] Step S2: Based on the preprocessed text to be identified, extract the sequence features, syntactic features and semantic features of the text, and construct the corresponding text feature maps according to the extracted features to obtain the feature map vectors;

[0048] Step S3: Based on the feature map vector of each text feature, a two-layer coarsening operation with residual connection is used to propagate the isomorphic node information within the graph, expand the update range of node information, update the context feature representation of the node, and obtain the node vector.

[0049] Step S4: Based on node vectors, an attention-based three-channel ensemble operation is used to aggregate the most relevant sequences, syntactic and semantic information of multiple graphs, enhance the heterogeneous node information between graphs, and obtain node enhancement vectors.

[0050] Step S5: Based on the node enhancement vector, perform pooling and fusion on the node enhancement vectors of the three text feature maps to obtain the text vector;

[0051] Step S6: Based on the text vector, the final classification is determined by a classifier to obtain the recognition result of the text to be recognized.

[0052] Specifically, in step S1, the first step is to acquire the self-media news text to be identified and preprocess it. Compared to traditional media such as television and newspapers, self-media content often lacks strict inspection and review by relevant departments, resulting in a lot of noisy data filled with typos, emoticons, and other inaccuracies. Therefore, as... Figure 2 As shown, the preprocessing of the self-media news text to be identified first involves data cleaning, which includes deleting redundant characters (i.e., text cleaning), word segmentation, stop word removal, stemming, and lowercase word correction. This process removes characters irrelevant to the propaganda identification task, resulting in clean text that facilitates dictionary construction and subsequent data processing. Secondly, the preprocessing also includes data augmentation, including synonym replacement, random insertion, random swapping, and random deletion, to address data imbalance and insufficient label data in the text. In this embodiment, the text preprocessing includes operations such as data cleaning, word segmentation, stop word removal, lowercase restoration, word simplification, and data augmentation.

[0053] The above preprocessing specifically includes the following steps:

[0054] Step S1.1: Clean the text to be identified by removing characters irrelevant to the propaganda identification task, including cleaning the text and removing special characters such as URL tags and punctuation marks; segmenting the text into words to obtain a token sequence; deleting stop words by removing words that appear frequently in the text but have little practical significance; converting all words in the corpus to lowercase; and stemming the words by using NLTK tools to perform lemmatization, simplifying word forms, and extracting word stems.

[0055] Step S1.2: Perform data augmentation on the cleaned text to be identified, including synonym replacement (randomly selecting a word in the sentence and replacing it with a synonym while maintaining the sentence's meaning); random insertion (randomly selecting a word in the sentence and randomly inserting its synonym); random swapping (randomly selecting two words in the sentence and swapping their positions); and random deletion (randomly selecting a word in the sentence and deleting it). These data augmentation operations address the issues of data imbalance and insufficient label data in the text.

[0056] In step S2, news propaganda can be spread through various propaganda techniques such as repetitive information, logical fallacies, and sentence transformation. Therefore, based on the preprocessed text to be identified, sequence features, syntactic features, and semantic features are extracted, and three text feature maps are constructed according to the extracted features, obtaining three feature map vectors. A text graph is a graph structure used to represent text data and the relationships between texts. In this embodiment, nodes in the text graph represent words in the text, and edges represent a certain relationship between words. Based on different relationships between nodes (sequence, syntax, and semantics), they are divided into different text feature maps, namely sequence feature maps, syntactic feature maps, and semantic feature maps. By extracting diverse dependencies between words, different types of propaganda techniques in the text are effectively mined. Through learning the features of propaganda news, the classification effect is made more accurate. The feature extraction steps include:

[0057] Step S2.1: Extract text sequence features. Using a sliding window mechanism, the co-occurrence information between any two words in the text, described by PMI, is utilized to generate word pair sequence weights. Specifically, the size of the sliding window is first defined as `window_size`. The text content is traversed, and the number of sliding windows N contained in the text is obtained based on the data length `seq_len`. windows Secondly, statistical analysis of the word "w" i w j The frequency of occurrence and word pairs in the corpus (w i ,w j The number of times a node exists in the corpus; finally, the edge weights between nodes in the sequence graph are defined using PMI. in

[0058] Step S2.2: Extract syntactic features from the text. Using Stanford NLP, syntactic features are extracted from the text. The frequency of syntactic dependencies between each word pair in the entire corpus is counted, generating word-pair syntactic weights. Specifically, firstly, the undirected relations between each word are extracted using the Stanford NLP parser. The extraction results are as follows: Figure 3 As shown; secondly, statistical word pairs (w i ,w j The number of times N has grammatical dependencies in the corpus ij Calculate the maximum and minimum frequency of word pairs with syntactic relations in the entire corpus. syn Min syn Finally, define the edge weights between nodes in the syntax graph.

[0059] Step S2.3: Extract semantic features from the text. Using BERT, extract semantic features of words, calculate the cosine similarity between two words, count the number of times each word pair has similar semantics in the entire corpus, and generate semantic weights for word pairs. The process of extracting semantic features from the text is as follows: Figure 4 As shown, first, the text is input into BERT to obtain its hidden layer vector h; second, word pairs (w) are calculated using the hidden layer vector h. i ,w j The cosine similarity between them is Next, a similarity threshold ρ is set. If the similarity value between two words exceeds the threshold ρ, it is determined that the two words have a semantic relationship in the current text. Then, word pairs (w) are counted. i ,w j The number of times N has grammatical dependencies in the corpus ij Calculate the maximum and minimum frequency of word pairs with syntactic relations in the entire corpus. sem Min sem Finally, define the edge weights between nodes in the syntax graph.

[0060] Step S2.4: Based on the extracted text sequence features, text syntax features, and text semantic features, construct the text sequence feature map G respectively. seq Text Syntactic Feature Map G syn and text semantic feature map G sem Each text feature map can be represented as G = (V, E), yielding three feature map vectors, where V (|V| = n) represents the set of nodes, and E (E ∈ {W}) represents the set of nodes. seq W syn W sem,}) describes the weight relationship between nodes.

[0061] In step S3, based on the feature map vector of each text feature, a two-layer coarsening operation with residual connections is used to propagate the isomorphic node information within the graph, expanding the update range of node information and updating the contextual feature representation of the nodes to obtain node vectors. By capturing the node relationships between different levels, the hidden propaganda techniques in the text are effectively mined, and covert propaganda news is identified. Step S3 is actually the propagation of isomorphic node information within the graph. First, a two-layer coarsening operation with residual connections is used to expand the update range of nodes. Then, a coordination layer is used to integrate global contextual information to obtain node vectors. Specifically, it includes the following steps:

[0062] Step S3.1: Construct a two-layer coarse-grained architecture for residual connections, such as... Figure 5 As shown, the architecture consists of two layers, with two symmetrical coarsening and thinning operation units within each layer. Coarsening and thinning operations are used to construct coarsening and thinning graphs, respectively. There are residual connections between the layers, representing the nodes of the lower-layer thinning graph and the upper-layer coarsening graph, which are used to update the node information of the upper layer.

[0063] Step S3.2: For each text feature map, neighborhood information is aggregated through graph convolution (i.e., graph convolutional neural network GCN) to capture the structural and attribute features of the text map, and the node representation H of the text map (hereinafter simply referred to as node representation or graph node representation) is obtained. Based on the graph node representation, the node representation of the three text feature maps is updated by applying the two-layer coarsening operation of residual connections to the constructed residual connection double-layer coarsening architecture. Specifically, the steps are as follows:

[0064] Step S3.2.1: Based on the original text feature map, structurally similar nodes are aggregated into supernodes through upper-level coarsening operations to generate an upper-level coarsened map, obtaining the node representation of the upper-level coarsened map. The node weight is defined as the number of nodes contained in the supernode. Specifically, if two nodes share the same neighbor set, they are considered structurally equivalent; therefore, structurally equivalent nodes are aggregated as supernodes. Then, in addition to the nodes sharing the same neighbor set, the node pair with the highest structural similarity is selected and aggregated into supernodes, node v. i and node v j The structural similarity between them is defined as: Among them W ij It is node v i and node v j The weights of the edges between nodes are defined, with D(·) representing the node weights. Finally, the remaining unselected nodes themselves form supernodes. A coarsening operation is then performed to aggregate nodes with similar structures into supernodes, generating a higher-level coarsened graph G′. cIts node representation matrix is ​​H′ c =M T H and the adjacency matrix are A′ c =M T HM, where M∈R n×m Let M be the transformation matrix, where n is the number of nodes in the original text graph and m is the number of nodes in the upper coarsened graph. The transformation matrix M is defined as follows:

[0065]

[0066] Where G represents the original graph, G′ c This represents the upper coarsening plot, v k Indicates a supernode.

[0067] Step S3.2.2: Based on the upper-level coarsened graph, aggregate the supernodes again through the lower-level coarsening operation to generate the lower-level coarsened graph G″. c Its node representation matrix is ​​denoted as H″ c The generated lower-level coarsened map is then refined to restore the original topology of the upper-level coarsened map, resulting in the lower-level refined map G″. r Obtain the node representation H″ of the lower-level refined graph. r Specifically, based on the node representation of the lower-level coarsened graph, the node representation of the current-level graph is calculated to obtain the refined text graph node representation H″. r =M′ T H″ c .

[0068] Step S3.2.3: After the node representations of the upper coarsened graph and the lower refined graph are residually connected, the original topological structure of the text graph is restored through the upper refined operation to generate the upper refined graph G′. r The node representation H′ of the upper-level refined graph is obtained. r Specifically, the upper-level coarsened graph node represents H′. c After step S3.1.2, a nonlinear transformation operation F(H′) is performed. c =Refine(Coarse(H′) c In this model, Coarse is the coarsening operation and Refine is the thinning operation; then, residual connections are established between layers, that is, the nodes of the upper coarsening graph are superimposed to represent H′. c and the text graph node representation F(H′) after nonlinear transformation c ), that is, H′ c +F(H′ c Finally, the text graph node representation H′ is obtained after refinement operations. r .

[0069] Step S3.2.4: The node representations of the original text graph and the upper-level refined graph are recovered by graph convolution after residual connection to obtain the updated node representation.

[0070] Step S3.3: Construct a coordination layer. The updated node representation is input into the graph convolutional network of the coordination layer for graph convolution operation, integrating global context information to obtain node vectors. Specifically, the coordination layer includes a graph convolutional neural network (GCN) with a softmax classifier, which takes the text graph node representation H′ after the last layer residual connection as input. c c∈{seq,syn,sem} is input into GCN, where node information is coordinated to obtain global context information, i.e., the text graph node vector is obtained.

[0071] In step S4, based on node vectors, an attention-based three-channel ensemble operation is used to aggregate the most relevant sequence, syntactic, and semantic information from multiple graphs, enhancing heterogeneous node information between graphs to obtain node enhancement vectors. While the two-layer coarsening operation using residual connections can update node representations within the text graph, sequence, syntactic, and semantic features are still separated from each other. Therefore, this embodiment employs attention-based three-channel feature ensemble, fusing other feature information within a specific relationship enhancement channel to enhance the representation of node information within the graph and the dependencies between words. This step, by coordinating sequence, syntactic, and semantic features from the three channels, reflects the differences in importance between different features, extracting complex and diverse propaganda techniques for identifying propaganda intent.

[0072] Specifically, such as Figure 6 As shown, three feature channels are constructed: sequence enhancement channel, syntactic enhancement channel, and semantic enhancement channel. Secondly, based on node vectors, and according to feature enhancement algorithms, enhanced representations of other features outside the current channel are obtained through node attention strategies and message passing mechanisms to enhance the features of the current channel. Finally, the feature enhancement algorithms within each channel are independent of each other, and multi-channel parallelization is achieved through matrix operations to obtain node enhancement vectors for the sequence, syntactic, and semantic feature maps. Based on the sequence enhancement channel, syntactic enhancement channel, and semantic enhancement channel, node enhancement vectors for the sequence, syntactic, and semantic feature maps are obtained according to feature enhancement algorithms, specifically including the following steps:

[0073] Taking sequence enhancement channels as an example, enhancing sequence features, such as... Figure 7 As shown, step S4.1 assigns different weights to the sequence relation graph based on the importance of the semantic relation graph and the syntactic relation graph;

[0074] Step S4.2: Based on the node attention strategy, the semantic and syntactic relations are fused into a single enhancement matrix;

[0075] Step S4.3: Based on the message passing mechanism, the sequence relationship and enhancement matrix are fused to enhance the sequence correlation representation of the text, and the node vector of the text sequence feature map is updated to obtain the node enhancement vector.

[0076] In this embodiment, firstly, the importance scores α of the semantic relation graph and the syntactic relation graph to the sequence relation graph are calculated respectively. c =softmax(FFN) c ([H seq H c ]),c∈{syn,sem}; Secondly, using the calculated importance scores, calculate the weighted sum O=α of semantic and syntactic relations. syn H syn +α sem H sem Finally, a gating mechanism is used to dynamically select semantic and syntactic relations, and these relations are fused with the sequence relation node representations to form a single augmentation matrix f = σ(W). seq ·(H seq ;O)+b seq Update the sequence text graph node vector H′. seq =f·H seq +(1-f)·O, at this point we get the node enhancement vector.

[0077] Similarly, syntactic and semantic feature enhancements are performed using similar methods. Syntactic feature enhancement includes: first, calculating the importance scores α of the sequence relation graph and the semantic relation graph to the syntactic relation graph, respectively. c =softmax(FFN) c ([H syn H c ]), c∈{seq,sem}; secondly, using the calculated importance scores, calculate the weighted sum O=α of sequence relations and semantic relations. seq H seq +α sem H sem Finally, a gating mechanism is used to dynamically select sequence relations and semantic relations, and these are fused with the syntactic relation node representations to form a single augmentation matrix f = σ(W). syn ·(H syn ;O)+b syn Update the syntactic text graph node vector H′. syn =f·H syn +(1-f)·O, at this point we get the node enhancement vector.

[0078] Semantic feature enhancement includes: First, calculating the importance scores α of the sequence relation graph and the syntactic relation graph to the semantic relation graph, respectively. c =softmax(FFN)c ([H sem H c ]), c∈{seq,syn}; secondly, using the calculated importance scores, calculate the weighted sum O=α of sequence relations and syntactic relations. seq H seq +α syn H syn Finally, a gating mechanism is used to dynamically select sequence relations and syntactic relations, and these are fused with the semantic relation node representations to form a single augmentation matrix f = σ(W). sem ·(H sem ;O)+b sem Update the semantic text graph node vector H′. sem =f·H sem +(1-f)·O, at this point we get the node enhancement vector.

[0079] In step S5, based on the node enhancement vectors, the node enhancement vectors of the three text feature maps are pooled and fused to obtain the text vector, which is: H final =Pooling(H′) seq ,H′ syn ,H′ sem ).

[0080] Finally, in step S6, based on the text vector, a classifier is used to make a final classification judgment to determine whether the text is persuasive or propaganda-oriented, and the text results are presented to the user in a visual way.

[0081] In the text intent recognition method proposed in this embodiment, a hierarchical feature graph network is constructed. Feature extraction, intra-graph information propagation, inter-graph information enhancement, and classification are performed sequentially within this network. Sequence features, semantic features, and syntactic features of the text are extracted, and a text feature graph is constructed based on these extracted features to obtain feature graph vectors. The constructed graph is analyzed to enhance the representation of word dependencies and to deeply uncover hidden propaganda techniques within the text. The hierarchical feature graph network is trained using a training set of texts labeled with propaganda intent. The trained hierarchical feature graph network is then used to identify news with hidden propaganda intent, improving the accuracy of propaganda intent recognition.

[0082] Example 2

[0083] This embodiment provides a text advertising intent recognition system based on hierarchical feature map networks, including:

[0084] The text acquisition and preprocessing module is used to acquire the text to be recognized and to preprocess the text to be recognized.

[0085] The feature vector construction module is used to extract sequence features, syntactic features, and semantic features of the preprocessed text to be identified, and to construct corresponding text feature maps based on the extracted features, thereby obtaining feature map vectors.

[0086] The information propagation module is used to propagate the isomorphic node information within the graph based on the feature map vector of each text feature using a two-layer coarsening operation with residual connections, thereby expanding the update range of node information, updating the context feature representation of the node, and obtaining the node vector.

[0087] The information enhancement module is used to aggregate the most relevant sequences, syntactic and semantic information of multiple graphs based on node vectors and an attention-based three-channel ensemble operation to enhance heterogeneous node information between graphs and obtain node enhancement vectors.

[0088] The text vector acquisition module is used to perform pooling and fusion of the node enhancement vectors of three text feature maps based on the node enhancement vectors to obtain the text vectors;

[0089] The classification module is used to perform the final classification judgment based on the text vector through a classifier, so as to obtain the recognition result of the text to be recognized.

[0090] Example 3

[0091] This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the processor executes the computer instructions, it completes the steps in the text propaganda intent recognition method based on hierarchical feature graph networks as described above.

[0092] Example 4

[0093] This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps in the text propaganda intent recognition method based on hierarchical feature map networks as described above.

[0094] The steps and methods involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0095] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0097] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for text propaganda intention recognition based on a hierarchical feature map network, characterized in that, The method comprises the following steps: obtaining to-be-recognized text, and preprocessing the to-be-recognized text; based on the preprocessed to-be-recognized text, extracting sequence features, syntactic features and semantic features of the text, and constructing corresponding text feature maps according to the extracted features to obtain feature map vectors; based on the feature map vectors of each text feature, propagating isomorphic node information within the graph by using a double-layer coarsening operation with residual connection to expand the update range of node information and update the context feature representation of the node to obtain a node vector; based on the node vector, using a three-channel integrated operation based on attention to aggregate the most relevant sequence, syntactic and semantic information of multiple graphs to enhance the isomorphic node information between graphs and obtain a node enhanced vector; based on the node enhanced vector, performing pooling fusion on the node enhanced vectors of the three text feature maps to obtain a text vector; based on the text vector, performing final classification determination through a classifier to obtain a recognition result of the to-be-recognized text; the method of propagating isomorphic node information within the graph by using a double-layer coarsening operation with residual connection to expand the update range of node information and update the context feature representation of the node to obtain a node vector based on the feature map vectors of each text feature comprises the following steps: constructing a double-layer coarsening architecture with residual connection; for each text feature map, aggregate neighborhood information through graph convolution to capture the structural features and attribute features of the text feature map, obtain node representation of the text graph, and based on the node representation, update the node representation of the three text feature maps by using the double-layer coarsening operation with residual connection through the constructed double-layer coarsening architecture with residual connection; construct a coordination layer, input the updated node representation into the graph convolution network of the coordination layer for graph convolution operation, integrate global context information, and obtain a node vector; the method of updating the node representation of the three text feature maps by using the double-layer coarsening operation with residual connection comprises the following steps: based on the original text graph, aggregate nodes with similar structures into super nodes through upper-layer coarsening operation to generate an upper-layer coarsening graph and obtain upper-layer coarsening graph node representation; based on the upper-layer coarsening graph, aggregate super nodes again through lower-layer coarsening operation to generate a lower-layer coarsening graph, the generated lower-layer coarsening graph restores the original topology structure of the upper-layer coarsening graph through thinning operation to generate a lower-layer thinning graph, and obtain lower-layer thinning graph node representation; after the node representations of the upper-layer coarsening graph and the lower-layer thinning graph are connected through residual connection, the original topology structure of the text graph is restored through upper-layer thinning operation to generate an upper-layer thinning graph, and upper-layer thinning graph node representation is obtained; after the node representations of the original text graph and the upper-layer thinning graph are connected through residual connection, the updated node representation is obtained through graph convolution.

2. The text propaganda intention recognition method based on the hierarchical feature map network according to claim 1, characterized in that, The preprocessing comprises: performing data cleaning on the to-be-recognized text, wherein the data cleaning comprises cleaning the text, word segmentation, deleting stop words, word case conversion and stem extraction; performing data enhancement on the to-be-recognized text after data cleaning, wherein the data enhancement comprises synonym replacement, random insertion, random exchange and random deletion. 3.The method of claim 1, wherein the method further comprises: determining a text propaganda intention of the text based on the hierarchical feature map network. the method of extracting sequence features, syntactic features and semantic features of the text comprises: The sequence feature of the text is extracted, and the sequence weight of the word pair is generated by using PMI to describe the co-occurrence information between any two words in the text based on a sliding window mechanism; The syntax feature of the text is extracted, and the syntax weight of the word pair is generated by counting the number of times that each pair of words has a syntax dependency in the entire corpus based on StanfordNLP; The semantic feature of the text is extracted, and the semantic weight of the word pair is generated by calculating the cosine similarity between the two words based on BERT and counting the number of times that each pair of words has similar semantics in the entire corpus. 4.The method of claim 1, wherein the method further comprises: determining a text propaganda intention of the text based on the hierarchical feature map network. The double-layer coarsening and refining architecture of the residual connection comprises two layers, each layer is provided with two symmetric coarsening operation units and refining operation units, a coarsening graph and a refining graph are respectively constructed through coarsening operation and refining operation, and a residual connection between the node representation of the lower-layer refining graph and the upper-layer coarsening graph is arranged between the layers to update the upper-layer node information.

5. The text propaganda intent recognition method based on the hierarchical feature map network according to claim 1, characterized in that, Based on the node vector, the most relevant sequence, syntax and semantic information of the multi-graph is aggregated by using the attention-based three-channel integrated operation to enhance the heterogeneous node information between the graphs, and the node enhanced vector is obtained, including: A sequence enhanced channel, a syntax enhanced channel and a semantic enhanced channel are constructed; Based on the node vector, the enhanced representation of other features outside the current channel is obtained through the node attention strategy and the message passing mechanism according to the feature enhancement algorithm, and is used to enhance the current channel feature; The feature enhancement algorithms in each channel are independently and parallelly run, and the node enhanced vectors of the sequence, syntax and semantic feature graphs are respectively obtained.

6. A text propaganda intention recognition system based on a hierarchical feature map network, characterized in that, The method for identifying the propaganda intention of the text based on the hierarchical feature graph network according to any one of claims 1-5 is implemented, including: A text acquisition and preprocessing module is configured to acquire the text to be identified and pre-process the text to be identified; A feature vector construction module is configured to extract the sequence feature, the syntax feature and the semantic feature of the text based on the pre-processed text to be identified, and to construct corresponding text feature graphs according to the extracted features and obtain feature graph vectors; An information propagation module is configured to propagate the isomorphic node information in the graph by using the double-layer coarsening and refining operation of the residual connection based on the feature graph vector of each text feature, expand the update range of the node information, update the context feature representation of the node, and obtain the node vector; An information enhancement module is configured to aggregate the most relevant sequence, syntax and semantic information of the multi-graph by using the attention-based three-channel integrated operation based on the node vector, enhance the heterogeneous node information between the graphs, and obtain the node enhanced vector; A text vector acquisition module is configured to pool and fuse the node enhanced vectors of the three text feature graphs based on the node enhanced vector, and obtain the text vector; A classification module is configured to perform final classification and determination by using a classifier based on the text vector, and obtain the identification result of the text to be identified.

7. An electronic device, characterized by comprising: The computer instructions stored on the memory and run on the processor complete the steps of the method for identifying the propaganda intention of the text based on the hierarchical feature graph network according to any one of claims 1-5 when the computer instructions are run by the processor.

8. A computer readable storage medium characterized by, A computer program product for storing computer instructions which, when executed by a processor, perform the steps of a method for text propaganda intent recognition based on a hierarchical feature map network according to any one of claims 1-5.

Images