A method and device for identifying false information on a trusted network

Abstract

The present invention discloses a method and device for identifying credible online false information, comprising the following steps: obtaining a first text to be detected, performing feature extraction on the first text to obtain deep semantic features and text element features; performing syntactic dependency analysis on the first text to obtain text style and extract syntactic dependency features; performing sentiment feature splicing on the first text to obtain sentiment feature features; performing feature splicing and multi-feature fusion on the deep semantic features and text element features, syntactic dependency features, and sentiment feature features of the first text to obtain fused features; inputting the fused features into a pre-trained online false information recognition model, and outputting a recognition result for the first text. The present invention improves the recognition performance of online text content and balances the accuracy and interpretability of the recognition model.

Description

Technical Field

[0001] The present invention relates to the field of text processing technology, and in particular to a method and device for identifying false information on a trusted network. Background Art

[0002] With the development of the internet, online disinformation is facing a governance dilemma: rapid growth in quantity, increasing variety, and increasing difficulty in identification. This is particularly true for online disinformation in the agricultural sector, which not only spreads misinformation but also hinders the normal production and sale of agricultural products. Therefore, optimizing online disinformation identification models is crucial for promoting the healthy exchange and dissemination of agricultural knowledge in cyberspace.

[0003] At present, most research on false information identification focuses on the improvement and optimization of identification models, and is committed to improving the accuracy and versatility of false information identification models. According to different technical methods, it can be divided into two types of identification methods: machine learning and deep learning.

[0004] Machine learning-based identification methods treat disinformation detection as a classification problem, improving recognition performance by refining feature extraction and classification models. Common disinformation features include content, sentiment, user, communication, and modality. Feature extraction methods commonly use methods like TF-IDF and named entity recognition (NER). Classification models are often stacked and upgraded around LR, RF, NB, KNN, SVM, and improved models. Limited by data size and model complexity, these methods suffer from poor generalization and transferability across different types of disinformation. Data acquisition and annotation are time-consuming and labor-intensive, and they are unable to capture deeper disinformation features.

[0005] Deep learning methods, developed based on machine learning, use deep neural networks to automatically learn the underlying deep representations of false information and extract more effective semantic features. These models are mostly based on and upgraded models such as Transformer, BERT, LSTM, CNN, and their improved counterparts. Because deep learning model training requires a large amount of data, its effectiveness on small sample datasets is limited. Furthermore, these methods exhibit logical black-box characteristics, and the models often lack interpretability. Therefore, recent research focuses on the credible identification of false information, including the credibility and interpretability of the identification results, and providing evidence for the interpretability of the false information identification results. However, these methods have overlooked the issue of balancing interpretability and accuracy in false information identification. Summary of the Invention

[0006] In order to solve the problem of difficulty in balancing the interpretability and accuracy of false information identification in the existing technology, the present invention proposes a method and device for identifying credible network false information, which aims to fully extract the deep semantic features and text element features of false information text, the text style and rhetorical features and emotional tendency features of syntactic dependency analysis, so as to comprehensively improve the recognition performance of the content of false information text on the Internet and balance its accuracy and interpretability.

[0007] The present invention discloses a method for identifying false information on a trusted network, which comprises:

[0008] Obtaining a first text to be detected, performing feature extraction on the first text, and obtaining deep semantic features and text element features;

[0009] Performing syntactic dependency analysis on the first text to obtain the text style and extract syntactic dependency features;

[0010] Performing sentiment tendency feature splicing on the first text to obtain sentiment tendency features;

[0011] Perform feature splicing and multi-feature fusion on the deep semantic features, text element features, syntactic dependency features, and sentiment tendency features of the first text to obtain fused features;

[0012] The fused features are input into a pre-trained network false information recognition model to output the false information recognition result of the first text.

[0013] Furthermore, obtaining the first text to be detected includes:

[0014] When identifying false information in text posted by users on a network platform, a first text to be detected is obtained by crawling;

[0015] The text element features include characters, geographical locations, events, materials, institutions, nature, behaviors and effects; the text styles include gossip style, curiosity style, commentary style, lyrical style and didactic style.

[0016] Furthermore, the feature extraction of the first text to obtain deep semantic features and text element features includes:

[0017] A word segmentation tool is used to segment the first text at the word level, and the RoBERTa pre-trained model is used to convert the first text after word segmentation into a word embedding vector. The word embedding vector extracted by fine-tuning the RoBERTa model is converted into a word embedding vector after Chinese word segmentation through the Feature Fusion layer to obtain text element features representing the structured information of the first text content.

[0018] Furthermore, the RoBERTa pre-training model is a bidirectional Transformer encoder based on the bidirectional Transformer model. Its core is composed of multiple Transformer-Encode modules. Each Encode module consists of a multi-head attention mechanism, residual connection and layer normalization, and a feedforward neural network.

[0019] Furthermore, the feature extraction of the first text to obtain deep semantic features and text element features specifically includes:

[0020] The word vector generation process uses a one-hot encoding representation to convert the words in the input text into a one-hot encoding form. At the same time, position encoding is introduced to represent the relative or absolute position of the word in the sequence. The two are superimposed to form a word embedding vector, which serves as the model input X.

[0021] The multi-head attention mechanism is a key component of the Transformer encoder. It uses the self-attention mechanism to perform parallel calculations on the input information, with multiple independent attention modules operating in parallel.

[0022] After the multi-head self-attention calculation, the multi-head attention mechanism passes the resulting hidden vector to the next stage twice: residual connection and layer normalization. Layer normalization converts the input into mean-variance, while residual connection superimposes the input X with the result of nonlinear transformation as the output. The input is then operated on by two fully connected layers of the feedforward neural network.

[0023] At the input layer, after each word embedding vector is encoded by the RoBERTa layer, a bidirectional association is established between the word embedding vectors, enabling the RoBERTa pre-trained model to learn the semantic features contained in each word embedding vector in different contexts;

[0024] The number of characters in each word in the sentence is counted to obtain a vector. For the word embedding vector output by the RoBERTa pre-trained model, the characters in the same word are averaged and the averaged vector is filled back into the original position.

[0025] Furthermore, performing syntactic dependency analysis on the first text to obtain text style and extract syntactic dependency features includes:

[0026] The RoBERTa-Feature Fusion layer is used to extract word embedding feature vectors of the false information text after word segmentation. DDParser is used to perform syntactic dependency analysis on the false information text sentences to identify the syntactic dependency relationships between entity words.

[0027] During syntactic dependency analysis, the result of Chinese word segmentation of the sentence is passed to DDParser;

[0028] The word-level vector obtained by the RoBERTa-Feature Fusion layer is fused with the syntactic dependency tree containing text rhetoric and style features obtained based on DDParser syntactic dependency analysis to form the final syntactic dependency feature.

[0029] Furthermore, the word-level vector obtained by the RoBERTa-Feature Fusion layer is fused with the syntactic dependency tree containing text rhetoric and style features obtained based on DDParser syntactic dependency analysis to form the final syntactic dependency features, including:

[0030] Use DDParser to analyze the syntactic dependency relationship of the target text and obtain the syntactic dependency tree processed by DDParser;

[0031] Replace all node words except Root in the syntactic dependency tree with the word embedding vectors (X1....X n ), n is the number of nodes in the dependency tree excluding the root node;

[0032] A two-dimensional dependency matrix is ​​formed according to the breadth and depth of the syntactic dependency tree, and the elements in the matrix are word vectors;

[0033] The constructed two-dimensional dependency matrix is ​​input into the Transformer layer by layer. The encoding layer encodes the input matrix and outputs the encoding matrix [[Y 1,1 ....Y 1,q ],...,[Y k,1 ....Y k,q ]], k is the number of rows in the dependency matrix, and q is the length of the sentence;

[0034] Aggregate information on the entire text rhetoric and style feature sequence and use average pooling processing Process the text sequence to form a syntactic dependency feature tensor [Z1....Z L ],Z n is the nth syntactic dependency feature.

[0035] Furthermore, the step of performing sentiment tendency feature splicing on the first text to obtain the sentiment tendency feature includes:

[0036] The false information text is converted into a word embedding vector through the RoBERTa-Feature Fusion layer, and the obtained word embedding vector is concatenated with the sentiment tendency label vector of the false information text to extract the sentiment tendency features of the first text.

[0037] Furthermore, the deep semantic features and text element features, syntactic dependency features, and sentiment tendency features of the first text are subjected to feature splicing and multi-feature fusion to obtain fused features, including:

[0038] Adjust the dimension of the sentiment information tensor from (batch_size, features) to (batch_size, 1, features) and convert it to floating point type;

[0039] The sentiment features, the text features of the underlying Transformer model, and the obtained syntactic dependency features are concatenated in the depth dimension to obtain the feature expression;

[0040] Transformer uses Positional Encoding to preserve the order information between words.

[0041] Furthermore, the step of inputting the fusion features into a pre-trained online false information recognition model and outputting a false information recognition result of the first text includes:

[0042] The fused features are linearly transformed, and the linearly transformed features are input into the ReLU activation function, which is then nonlinearly transformed to obtain the final features; the final features are input into a fully connected layer and mapped to the output space; the output is normalized through the softmax function to obtain the probability distribution of two categories to achieve the false information binary classification task of the first text; and the false information recognition result of the first text is output.

[0043] The present invention also discloses a trusted network false information identification device, which is used to implement any of the trusted network false information identification methods described above, and comprises:

[0044] A first acquisition module is used to acquire a first text to be detected, perform feature extraction on the first text, and obtain deep semantic features and text element features;

[0045] A second acquisition module is used to perform syntactic dependency analysis on the first text, obtain the text style and extract syntactic dependency features;

[0046] A third acquisition module is used to perform sentiment tendency feature splicing on the first text to obtain sentiment tendency features;

[0047] A feature fusion module is used to perform feature splicing and multi-feature fusion on the deep semantic features and text element features, syntactic dependency features, and sentiment tendency features of the first text to obtain fused features;

[0048] The false information identification module is used to input the fusion features into a pre-trained network false information identification model and output the false information identification result of the first text.

[0049] Due to the adoption of the above technical solution, the present invention has the following advantages: the present invention realizes the deep understanding and credible identification of false information texts, and the model integrates the feature interpretation of machine learning, the feature extraction of deep learning, and the trust factor representation of credible learning, which significantly improves the accuracy and deep understanding ability of false information identification, balances the interpretability and accuracy of false information identification, and provides a new theoretical basis and practical method for the credible identification of false information on the Internet. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments described in the embodiments of the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0051] Figure 1 It is a flow chart of a method for identifying false information on a trusted network according to the present invention;

[0052] Figure 2 This is the encoder-decoder structure diagram of the Transformer of the present invention;

[0053] Figure 3 This is a structural diagram of the RoBERTa-Feature Fusion model of the present invention;

[0054] Figure 4 is an example diagram of the syntactic dependency tree of the present invention;

[0055] Figure 5 It is a logical structure diagram of the syntactic dependency feature extraction of the present invention;

[0056] Figure 6 It is a structural diagram of the network false information detection device of the present invention. DETAILED DESCRIPTION

[0057] The present invention will be further described with reference to the accompanying drawings and embodiments. The embodiments described are only a part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by those skilled in the art should fall within the scope of protection of the embodiments of the present invention.

[0058] The present invention fully utilizes the powerful feature learning ability of the text pre-training model in deep learning and the multi-granularity text element representation ability of the feature fusion layer through the RoBERTa-Feature Fusion (RoBERTa+feature fusion pre-trained language model) layer, and integrates the RoBERTa-Feature Fusion layer with the Chinese dependency syntax analysis tool DDParser (BaiduDependency Parser), focusing on the syntactic dependency characteristics of false information text, capturing the style and rhetorical features of false information text, combining machine learning to extract the emotional tendency features of false information text, and using the Transformer deep learning model based on the self-attention mechanism to fuse and encode multi-dimensional features, thereby improving the recognition performance of the content of false information text on the Internet and balancing the accuracy and interpretability of the recognition model.

[0059] The technical solutions provided by the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0060] Figure 1 This is a flow chart of a trusted network false information identification method based on the RFEDTransformer (RoBERTa-Feature Fusion-Emotion-DDParser-Transformer) integrated algorithm of the present invention. Figure 1 The method for identifying false information on the Internet is described as follows.

[0061] S102: Obtain a first text to be detected on a network platform;

[0062] When identifying false information in texts posted by users on a network platform, a first text to be detected is obtained by a python crawler.

[0063] S104: Extract features from the first text to obtain deep semantic features and text element features;

[0064] After obtaining the first text, use the word segmentation tool to segment the first text at the word level, use the RoBERTa pre-trained model to convert the first text after word segmentation into a word embedding vector, and use the Feature Fusion layer to convert the word embedding vector extracted by fine-tuning the RoBERTa model into a word embedding vector after Chinese word segmentation to obtain text element features such as people, geographic locations, and events that represent the structured information of the first text content.

[0065] In the embodiment of the present invention, when the text element features are extracted in S104, the RoBERTa model is a bidirectional Transformer encoder based on the bidirectional Transformer model, and its core is composed of multiple Transformer-Encode modules. Each Encode module is composed of a multi-head attention mechanism, residual connection and layer normalization, and a feedforward neural network. The model structure is shown in Figure 2 .

[0066] S104.1: The word embedding process uses one-hot encoding to convert the words in the input text into a one-hot encoding. Positional encoding is also introduced to represent the relative or absolute position of a word in the sequence. These two are combined to form the word embedding vector, which serves as the model input X.

[0067] S104.2: The multi-head attention mechanism is a key component of the Transformer encoder. It uses the self-attention mechanism to perform parallel calculations on the input information. Multiple independent attention modules operate in parallel. Its core operation formula is shown in Equation (1).

[0068]

[0069] Among them, Attention is the attention mechanism function, Q, K, V are the query matrix, key matrix, value matrix, d k Represents the dimension of the column vector in K, and Softmax is a normalized exponential function. After the multi-head self-attention calculation, the multi-head attention mechanism delivers the obtained hidden vector twice to the next link: residual connection and layer normalization. Layer normalization converts the input into mean variance, and residual connection superimposes the input X with the result obtained by nonlinear transformation as the output. The input is then operated on by two fully connected layers of the feedforward neural network, using the formula shown in (2):

[0070] W e =max(0,XW0+b0)W′0+b′ (2)

[0071] Among them, {W e ,W0′} is the weight matrix of the two connection layers, and {b0,b′} is the bias term of the two connection layers.

[0072] S104.3: At the input layer, after each word embedding vector is encoded by the RoBERTa layer, a bidirectional association is established between the word embedding vectors, enabling the model to learn the semantic features contained in each word embedding vector in different contexts. In this example, the semantic features of the false information text "Genetically modified soybeans can cause cancer" and the corresponding true information text "No pharmacological or clinical studies have shown a link between genetically modified soybeans and cancer" differ significantly. The false information text's assertions are scientifically unfounded, while the true information text objectively states the facts and provides support from pharmacological and clinical research.

[0073] S104.4: In the Feature Fusion layer, the example uses the Jieba vocabulary to segment the text. For example, the sentence "Genetically modified soybeans can cause cancer" is segmented into ['genetically modified', 'soybean', 'will', 'cause', 'cancer']. The number of characters in each word in the sentence is counted to obtain a vector of [3, 2, 1, 2, 2]. For the word embedding vector output by RoBERTa, the characters in the same word are averaged and then these processed vectors are filled back into the original position. The details of the logical structure are shown in [1]. Figure 3 .

[0074] S106: Perform syntactic dependency analysis on the first text to obtain text style and extract syntactic dependency features.

[0075] After obtaining the deep semantic features and text element features of the first text, DDParser is used to perform syntactic dependency analysis on the first text, and RoBERTa-Feature Fusion is used to represent the word vectors in the syntactic dependency tree. This method systematically captures the syntactic relationship between words in the false information text and extracts the text style and syntactic dependency features of the first text.

[0076] When S106 extracts text style and rhetorical features, the embodiment performs syntactic dependency analysis on the sentence-level text of the false information and constructs a syntactic dependency tree through DDParser, uses the word segmentation vectors extracted by the RoBERTa-Feature Fusion layer to replace the non-root nodes with the word vectors corresponding to the node words, and performs matrix processing based on the tree structure. After processing through the multi-layer self-attention mechanism and feedforward neural network layer of the Transformer, the feature representation of each position is finally obtained, thereby obtaining the style and rhetorical features of the false information text.

[0077] S106.1: The word embedding feature vectors of the false information text after Jieba segmentation were extracted using the RoBERTa-Feature Fusion layer. DDParser was used to perform syntactic dependency analysis on the false information text sentences, identifying 14 types of syntactic dependencies between entity words, including subject-verb (SBV), verb-object (VOB), preposition-object (POB), adverbial-verbal (ADV), verb-complement (CMP), attributive-mediant (ATT), positional (F), coordination (COO), concurrent verb (DBL), double object (DOB), continuous verb (VV), clause (IC), function word (MT), and heart (HED).

[0078] S106.2: During syntactic dependency analysis, the result of Chinese word segmentation of the sentence is passed to DDParser, where W, H, and D are the word list (wordList), the head node list (HeadList), and the dependency list (DeprelList) respectively. Example "Genetically modified soybeans can cause cancer": {'word': ['genetically modified', 'soybean', 'will', 'lead to', 'cancer'], 'head': [2, 4, 4, 0, 4], 'deprel': ['ATT', 'SBV', 'ADV', 'HED', 'VOB']}; the corresponding real information example "Pharmacological and clinical studies have not shown that genetically modified soybeans are related to cancer": {'word': ['pharmacological', 'and', 'clinical','on','no','research','shows','GM','soybean''is','cancer','related'],'head':[6,4,4,1,6,7,0,12,12,12,10,7],'deprel':['SBV','MT','ATT','COO','ADV','IC','HED','ADV','SBV','ADV','POB','VOB']}. See the syntactic dependency tree for the example. Figure 4 .

[0079] S106.3: The word-level vector obtained by the RoBERTa-Feature Fusion layer is fused with the syntactic dependency tree containing text rhetoric and style features obtained based on DDParser syntactic dependency analysis to form the final syntactic dependency feature. Figure 4 The specific process is as follows:

[0080] S106.3.1: Use DDParser to analyze the syntactic dependency relationship of the target text (s) and obtain the syntactic dependency tree (t) processed by DDParser.

[0081] S106.3.2: Replace all node words except Root in the syntactic dependency tree (t) with the word embedding vectors (X1....X n ), n is the number of nodes in the dependency tree excluding the root node.

[0082] S106.3.3: A two-dimensional dependency matrix is ​​constructed based on the width and depth of the syntactic dependency tree. The elements in the matrix are word vectors. If the sentence length exceeds a fixed length L, the excess is truncated, and the insufficient part is padded with zero vectors. Assume that the unit of the fixed length L is a word. The following example uses a fixed length L of 10 as an example.

[0083] S106.3.4: Input the constructed two-dimensional dependency matrix into the Transformer layer by layer. The encoding layer encodes the input matrix and outputs the encoding matrix [[Y 1, 1....Y 1,10 ],...,[Y k,1 ....Y k,10 ]], k is the number of rows in the dependency matrix.

[0084] S106.3.5: Aggregate information on the entire text rhetoric and style feature sequence using average pooling Process the text sequence to form a syntactic dependency feature tensor [Z1....Z 10 ],Z n is the nth syntactic dependency feature, where n ranges from 1 to 10.

[0085] S108: Perform sentiment tendency feature splicing on the first text to obtain sentiment tendency features.

[0086] After obtaining the deep semantic features, text element features, and syntactic dependency features of the first text, when extracting the sentiment tendency features in S108, the embodiment converts the false information text into a word embedding vector through the RoBERTa-Feature Fusion layer, and then splices the obtained embedding vector with the sentiment tendency label vector of the false information text to extract the sentiment tendency features of the first text. The sentiment tendency feature splicing model structure is shown in FIG. Figure 5 .

[0087] S110: performing feature splicing and multi-feature fusion on the deep semantic features and text element features, syntactic dependency features and sentiment tendency features of the first text to obtain fused features.

[0088] After obtaining the deep semantic features, text element features, syntactic dependency features, and sentiment tendency features of the first text, the text vectors of the above features after being encoded by RoBERTa are spliced ​​together, and the features output by different models are fused and encoded using the Transformer model based on the self-attention mechanism to obtain fused features.

[0089] When S110 performs multi-feature fusion, the deep semantic features and text element features, sentiment features, and syntactic dependency features of the first text are concatenated with the text vectors encoded by RoBERTa, and multi-feature fusion is performed through Transformer. The feature concatenation process is as follows:

[0090] S110.1: Adjust the dimension of the sentiment information tensor from (batch_size, features) to (batch_size, 1, features) and convert it to floating point type. batch_size is the batch size parameter.

[0091] S110.2: The sentiment features in S110.1, the text features of the underlying Transformer model, and the syntactic dependency features obtained in S106 are combined in the depth dimension to obtain a richer feature expression, as shown in formula (3):

[0092]

[0093] Among them, emo_feats represents emotional features, trsf_feats represents features extracted from the underlying Transformer model, and ddp_feats represents syntactic dependency features.

[0094] The Transformer in the example uses Positional Encoding (formulas (4) and (5) below) to preserve the order information between words.

[0095]

[0096] Among them, t is the actual position of the token in the sequence, PE t ∈R f is the position vector of the song token, is the i-th element of this position vector, d mofel It is the dimension of this token.

[0097] S112: Input the fused features into a pre-trained online false information recognition model, and output a false information recognition result of the first text.

[0098] After obtaining the fused features, the fused features are input into Transformers to complete the downstream false information classification task:

[0099] (1) Perform a linear transformation on the fused features and convert the dimension of the fused features from the original size to 256.

[0100] (2) The linearly transformed features are input into the ReLU activation function, which is then nonlinearly transformed to obtain the final feature representation last_feature.

[0101] (3) The feature last_feature is input into a fully connected layer (fc) to map it to the output space (dimension is 2).

[0102] (4) The output is normalized by the softmax function to obtain the probability distribution of two categories to achieve the false information binary classification task of the first text.

[0103] Output the false information identification result of the first text.

[0104] In this example, a dataset of agricultural misinformation was constructed, containing 3,572 pieces of both false and true agricultural information, for training and validating the RFED Transformer model. The data was sourced from mainstream Chinese rumor-busting websites: the Internet Rumor Exposure Platform (a platform within the China Internet Joint Rumor-busting Platform), Today's Rumor Refutation (a platform within the China Internet Joint Rumor-busting Platform), Tencent News's Rumor-busting Platform, Baidu's Rumor-busting Platform, the Science Rumor-busting Platform, and the Weibo Community Management Center.

[0105] After completing the construction of the false information text dataset, the embodiment designed a comparative test and an ablation experiment to verify the feasibility and effectiveness of the false information identification method proposed in the present invention.

[0106] Comparative Experiments: To verify the effectiveness of the proposed RFEDTransformer model and reasonably evaluate the experimental results, experiments were conducted on the online false information dataset constructed by this invention in the same experimental environment, compared with other baseline methods. The comparative experimental results are shown in Table 1. The proposed RFEDTransformer model achieved the highest scores in all evaluation metrics, with an accuracy of 0.9563, a recall of 0.9565, and an F1 value of 0.9563, indicating that the model has the best overall performance in the task of identifying online false information text.

[0107] Table 1 Results of online false information identification

[0108]

[0109] Ablation experiment: In order to verify the contribution and influence of each module in the RFEDTransformer model proposed in this invention on the overall performance of the model, and to evaluate the effectiveness and necessity of each module, based on the baseline model RoBERTa, this invention designed 4 groups of ablation experiments for verification.

[0110] Ablation experiment 1: Remove the Emotion module from the model. This removes the sentiment labels from the false information and does not incorporate sentiment features into the model. The model only represents the fusion of semantic features, Chinese word fusion features, and syntactic dependency features.

[0111] Ablation Experiment 2: Remove the DDParser module from the model. Syntactic analysis of the false information text is no longer performed, and contextual features such as style and rhetoric are not extracted from the model. Instead, the model focuses on extracting semantic features and Chinese word fusion features, and combining them with sentiment features for a fusion representation.

[0112] Ablation experiment 3: Remove the Feature Fusion module from the model. Remove the lexical feature extraction for Chinese word segmentation and word fusion from the model, and only use the semantic features of word vectors extracted by RoBERTa, combined with syntactic dependency features and sentiment features for fusion representation.

[0113] Ablation experiment 4: The Transformer module is removed from the model. The model extracts the semantic features of the false information text, Chinese word fusion features, syntactic dependency features, and sentiment feature vectors, concatenating them and directly inputting them into a fully connected layer with a softmax function for false information identification and classification.

[0114] Table 2 Ablation experiment results

[0115]

[0116] The experimental results are shown in Table 2. The index results of the four groups of ablation experiments are all lower than those of the model proposed in this invention. Removing any module of the model will lead to a decrease in model performance, which shows that all four modules are effective for the experiment.

[0117] The present invention also provides an embodiment of a trusted network false information identification device, which is used to implement the trusted network false information identification method described in the above embodiment, and includes:

[0118] A first acquisition module is used to acquire a first text to be detected, perform feature extraction on the first text, and obtain deep semantic features and text element features;

[0119] A second acquisition module is used to perform syntactic dependency analysis on the first text, obtain the text style and extract syntactic dependency features;

[0120] A third acquisition module is used to perform sentiment tendency feature splicing on the first text to obtain sentiment tendency features;

[0121] A feature fusion module is used to perform feature splicing and multi-feature fusion on the deep semantic features and text element features, syntactic dependency features, and sentiment tendency features of the first text to obtain fused features;

[0122] The false information identification module is used to input the fusion features into a pre-trained network false information identification model and output the false information identification result of the first text.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for identifying false information on a trusted network, characterized in that: include: Obtaining a first text to be detected, performing feature extraction on the first text, and obtaining deep semantic features and text element features; Performing syntactic dependency analysis on the first text, obtaining a syntactic dependency tree and extracting syntactic dependency features; Extracting word embedding vectors and sentiment label vectors from the first text, and concatenating the word embedding vectors and sentiment label vectors to obtain sentiment features of the first text; Perform feature splicing and multi-feature fusion on the deep semantic features, text element features, syntactic dependency features, and sentiment tendency features of the first text to obtain fused features; Inputting the fused features into a pre-trained network false information recognition model and outputting a false information recognition result of the first text; The performing syntactic dependency analysis on the first text, obtaining a syntactic dependency tree and extracting syntactic dependency features includes: Extract the word embedding feature vector of the false information text after word segmentation through the RoBERTa-Feature Fusion layer; DDParser is used to perform syntactic dependency analysis on false information text sentences to identify the syntactic dependency relationships between entity words; During syntactic dependency analysis, the result of Chinese word segmentation of the sentence is passed to DDParser; The word embedding feature vector obtained by the RoBERTa-Feature Fusion layer is fused with the syntactic dependency tree containing text rhetoric and style features obtained based on DDParser syntactic dependency analysis to form the final syntactic dependency feature.

2. The method for identifying false information on a trusted network according to claim 1, wherein: The feature extraction of the first text to obtain deep semantic features and text element features includes: A word segmentation tool is used to segment the first text at the word level, and the RoBERTa pre-trained model is used to convert the first text after word segmentation into a word embedding vector. The word embedding vector extracted by fine-tuning the RoBERTa model is converted into a word embedding vector after Chinese word segmentation through the Feature Fusion layer to obtain text element features representing the structured information of the first text content.

3. The method for identifying false information on a trusted network according to claim 2, wherein: The RoBERTa pre-training model is a bidirectional Transformer encoder based on the bidirectional Transformer model. Its core is composed of multiple Transformer-Encode modules. Each Encode module consists of a multi-head attention mechanism, residual connection and layer normalization, and a feedforward neural network.

4. The method for identifying false information on a trusted network according to claim 3, wherein: The feature extraction of the first text to obtain deep semantic features and text element features specifically includes: The word embedding process uses a one-hot encoding method to convert the words in the first text into a one-hot encoding form. At the same time, position encoding is introduced to represent the relative or absolute position of the word in the sequence. The two are superimposed to form a word embedding vector, which serves as the input of the online false information detection model. The multi-head attention mechanism is a key component of the Transformer encoder. It uses the self-attention mechanism to perform parallel calculations on the input information, with multiple independent attention modules operating in parallel. After the multi-head self-attention calculation, the multi-head attention mechanism passes the resulting hidden vector to the next stage twice: residual connection and layer normalization. Layer normalization converts the input into mean-variance, and residual connection superimposes the input with the result of nonlinear transformation as the output. The input is then operated on by two fully connected layers of the feedforward neural network. At the input layer, after each word embedding vector is encoded by the RoBERTa layer, a bidirectional association is established between the encoded word embedding vectors, enabling the RoBERTa pre-trained model to learn the semantic features contained in each word embedding vector in different contexts; The number of characters in each word in the sentence is counted to obtain a vector. For the word embedding vector output by the RoBERTa pre-trained model, the characters in the same word are averaged and the averaged vector is filled back into the original position.

5. The method for identifying false information on a trusted network according to claim 1, wherein: The word embedding feature vector obtained by the RoBERTa-Feature Fusion layer is fused with the syntactic dependency tree containing text rhetoric and style features obtained based on DDParser syntactic dependency analysis to form the final syntactic dependency features, including: Analyze the syntactic dependency relationship of the first text using DDParser to obtain a syntactic dependency tree processed by DDParser; The root node is a word node in the syntactic dependency tree that does not depend on other words. All node words in the syntactic dependency tree except the root node are replaced with the word embedding vector (X1....X n′ ), n ′ is the number of nodes in the dependency tree excluding the root node; A two-dimensional dependency matrix is ​​formed according to the breadth and depth of the syntactic dependency tree, and the elements in the matrix are word vectors; The constructed two-dimensional dependency matrix is ​​input into the Transformer layer by layer. The encoding layer encodes the input matrix and outputs the encoding matrix [[Y 1,1 ....Y 1,n ],...,[Y k, 1....Y k,n ]], k is the number of rows in the dependency matrix, and n is the length of the sentence; Aggregate information on the entire text rhetoric and style feature sequence and use average pooling processing Process the text sequence to form a syntactic dependency feature tensor [Z1....Z q ],Z j is the jth syntactic dependency feature.

6. The method for identifying false information on a trusted network according to claim 1, wherein: The extracting of word embedding vectors and sentiment label vectors from the first text, and concatenating the word embedding vectors and sentiment label vectors to obtain sentiment features of the first text includes: The false information text is converted into a word embedding vector through the RoBERTa-Feature Fusion layer, and the obtained word embedding vector is concatenated with the sentiment tendency label vector of the false information text to extract the sentiment tendency features of the first text.

7. The method for identifying false information on a trusted network according to claim 1, wherein: The step of performing feature splicing and multi-feature fusion on the deep semantic features, text element features, syntactic dependency features, and sentiment tendency features of the first text to obtain fused features includes: Adjust the dimension of the sentiment feature tensor from (batch_size, features) to (batch_size, 1, features) and convert it to a floating point type. The sentiment tendency features, text element features and the obtained syntactic dependency features are spliced ​​in the depth dimension to obtain the feature expression; Transformer uses Positional Encoding to preserve the order information between words.

8. The method for identifying false information on a trusted network according to claim 1, wherein: The step of inputting the fusion features into a pre-trained online false information recognition model and outputting a false information recognition result of the first text includes: The fused features are linearly transformed, and the linearly transformed features are input into the ReLU activation function, which is then nonlinearly transformed to obtain the final features; the final features are input into a fully connected layer and mapped to the output space; the output is normalized through the softmax function to obtain the probability distribution of two categories to achieve the false information binary classification task of the first text; and the false information recognition result of the first text is output.

9. A trusted network false information identification device, used to implement the trusted network false information identification method according to any one of claims 1 to 8, characterized in that: include: A first acquisition module is used to acquire a first text to be detected, perform feature extraction on the first text, and obtain deep semantic features and text element features; A second acquisition module is used to perform syntactic dependency analysis on the first text, obtain a syntactic dependency tree and extract syntactic dependency features; A third acquisition module is used to extract word embedding vectors and sentiment label vectors from the first text, and concatenate the word embedding vectors and sentiment label vectors to obtain sentiment features of the first text; A feature fusion module is used to perform feature splicing and multi-feature fusion on the deep semantic features and text element features, syntactic dependency features, and sentiment tendency features of the first text to obtain fused features; The false information identification module is used to input the fusion features into a pre-trained network false information identification model and output the false information identification result of the first text.

Images