Single-mode rumor detection method with automatic rumor refuting ability

Abstract

The invention discloses a single-mode rumor detection method with automatic rumor refuting ability, and relates to intelligent rumor detection. Causal relationship triples are extracted from a rumor text in a fine-grained mode, and modeling is conducted on the causal relationship triples in a causal map mode; extracting embedded expression features of the causal atlas through the graph convolutional neural network; utilizing a text feature extractor to obtain features of a rumor text, and fusing the features with features of the causal atlas to form causal enhanced single-mode rumor features; a classifier with an attention mechanism is utilized to classify the single-mode features after causal enhancement, and rumor detection is completed; and taking rumor text data as input, and generating corresponding rumor refuting text content for the rumor text data through the rumor refuting text generator substructure part in the model overall framework. The rumor detection performance is improved by utilizing the causal relationship of the rumor text, the rumor refuting text generation is introduced for the model, so that the model has the automatic rumor refuting ability, the performance of single-mode rumor detection is better, and meanwhile, the effect of automatically generating the rumor refuting text is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent rumor detection, and in particular to a single-mode rumor detection method with automatic rumor refuting capability. Background Art

[0002] "Rumors stop at the wise." For potential rumor recipients with non-malicious tendencies, when they can understand the truth behind the rumor and correct their cognitive biases, they will naturally avoid the negative impact that rumors may bring. Therefore, publishing rumor-refuting texts has always been one of the most common and core links in the rumor governance process.

[0003] At present, in order to curb the spread of rumors, some public welfare rumor-refuting platforms have also emerged in China. Help people to learn about the scientific knowledge and truth of the epidemic in a timely manner. Rumor-refuting texts have received widespread attention and application due to their significant significance in governing rumors. Tu et al. (Yan Tu, Linqi Cheng, Liyi Liu, Zongmin Li, Benjamin Lev. How to improve the rumor-confutation ability of official rumor refuting account on social media: A Chinese case study [J]. Computer Communications, 2023, Volume 209, Pages 331-348.) discussed the rumor-refuting capabilities and differences of various official rumor-refuting accounts in China and provided relevant suggestions. However, the current rumor-refuting platforms mainly construct rumor-refuting texts through manual writing and manually publish them. This method has significant defects in efficiency and timeliness. Zhang Weidong et al. (Zhang Weidong, Li Songtao, Mao Xiumei. The influence of rumor-debunking information on the formation of social media opinion leaders from the perspective of role evolution - based on the law of public opinion leadership [J]. Intelligence Theory and Practice, 2024, 47(1): 110-119) proposed that factors including rumor-debunking information will affect the formation of "opinion leaders" in the process of rumor propagation. They proposed that the quality of rumor-debunking text will directly affect the user's decision to participate in rumor-debunking. The actual performance of rumor-debunking text in the rumor-debunking process is a widely concerned issue. Zhang Ran et al. (Zhang Ran, Xia Zhijie. Research on the synergistic influence of social media rumor-debunking information text features on rumor-debunking effect [J]. Science and Management, 2023, 43(2): 72-78) explored the influence of rumor-debunking text features (content features, form features) on rumor-debunking effect. They believe that rumor-debunking information should pay attention to the comprehensiveness of content and the way of information presentation. Li Chen (Li Chen. Research on rumor-debunking text generation technology with knowledge introduction [D]. Xiamen University, 2021) pointed out in his master's thesis that the knowledge-driven and constrained approach can be used to automatically generate rumor-debunking texts, but he did not further consider how to use it in rumor detection. In general, in the existing work, the research on rumor-debunking texts lacks the ability to automatically refute rumors while detecting them, and focusing on the automatic generation of rumor-debunking texts is one of the research directions that needs to be explored urgently. The combination of rumor detection and automatic generation of rumor-debunking texts is a very practical and valuable work. Summary of the invention

[0004] The purpose of the present invention is to provide a single-modal rumor detection method with automatic rumor refuting capability to solve the problem that the single-modal rumor detection method does not have the ability to automatically refute rumors. The invention aims to use the causal relationship of rumor text to improve the rumor detection performance, and introduce rumor refuting text generation into the model so that the model has the ability to automatically refute rumors, thereby achieving better performance in single-modal rumor detection and the effect of automatically generating rumor refuting text. The present invention enables the model to have the ability to automatically refute rumors while completing rumor detection.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The single-mode rumor detection method does not have the ability to automatically refute rumors, including the following steps:

[0007] 1) Extract the causal triplets contained in the rumor text in a fine-grained manner and model them in the form of a causal graph;

[0008] 2) Extracting embedded expression features of causal graphs through graph convolutional neural networks;

[0009] 3) Use the text feature extractor to obtain the features of the rumor text, and integrate them with the features of the causal graph to form a causally enhanced single-modal rumor feature;

[0010] 4) Using a classifier with an attention mechanism to classify the causally enhanced unimodal features to complete rumor detection;

[0011] 5) Take the rumor text data as input and pass it through the rumor-refuting text generator substructure in the overall model framework to generate corresponding rumor-refuting text content.

[0012] In step 1), the specific method of extracting the causal relationship triples contained in the rumor text in a fine-grained manner and modeling them in the form of a causal graph is:

[0013] 1.1) Preprocessing: Perform word segmentation and part-of-speech tagging on rumor text;

[0014] 1.2) Named Entity Recognition: Named entity recognition technology (such as BERT-based NER) is used to extract entities in the text, such as names of people, places, and organizations. These entities will serve as potential causal nodes.

[0015] 1.3) Causal relationship mining: Use the causal relationship algorithm (fast greedy equivalent search algorithm) to analyze sentences and paragraphs in the text and extract causal relationship triplets (cause-relationship-effect);

[0016] 1.4) Construct a causal graph: Based on the extracted causal triples, a rumor-related causal graph is constructed. The nodes in the graph are entities or events, and the edges are the causal relationships between them.

[0017] In step 2), the specific method of extracting the embedded expression features of the causal graph through the graph convolutional neural network is:

[0018] 2.1) Constructing the adjacency matrix: Based on the structural information of the causal graph, the adjacency matrix A is obtained from the graph. The elements in the matrix represent the causal relationship between nodes.

[0019] 2.2) Obtain node feature matrix: For each node (entity or event), obtain the feature matrix of the node set V of the graph through the word embedding algorithm In the formula, The Glove algorithm is used to obtain the corresponding word vector feature of the node v.

[0020] 2.3) Graph Convolutional Neural Network: Taking the adjacency matrix A and the node feature matrix H as input, the graph convolutional neural network (GCN) is used to aggregate information and extract features to obtain the embedded feature representation of each node, that is, where f gcn () represents the information aggregation and feature extraction process of graph convolutional neural network.

[0021] In step 3), the specific method of using a text feature extractor to obtain the features of the rumor text, and fusing them with the features of the causal graph to form a causally enhanced unimodal rumor feature is:

[0022] 3.1) Text feature extraction: Use the pre-trained text feature extractor Text2Vec to encode the rumor text and obtain the feature vector of the rumor text;

[0023] 3.2) Feature fusion: The feature vector of the rumor text is fused and spliced ​​with the embedded expression features of the causal graph obtained in step 2) to achieve causal semantic enhancement of the rumor text features and form rumor text features enhanced with causal semantics. This feature contains both text content information and causal relationship information.

[0024] In step 4), the specific method of using a classifier with an attention mechanism to classify the causally enhanced unimodal features to complete rumor detection is:

[0025] 4.1) Attention Mechanism: The self-attention mechanism in the Transformer model or a similar attention model is used to calculate the attention weight parameters between features through the query matrix, key matrix, and value matrix of the features to enhance the model's attention to the key semantic content in the text features after causal enhancement;

[0026] 4.2) Classification: Combined with the attention weight parameter, the attention-weighted features are input into the softmax classifier to complete the classification of rumor text features. The classification is performed and the result of rumor classification is output, that is, whether the text content is a rumor.

[0027] In step 5), the rumor text data is taken as input, and the rumor refuting text generator substructure part in the overall model framework is used to generate the corresponding rumor refuting text content. The specific method is:

[0028] 5.1) Model selection: According to the characteristics of the rumor-debunking text generation task, select or train a suitable sequence-to-sequence model, such as BERT, T5, or ChatGLM, as the rumor-debunking text generator to generate rumor-debunking text;

[0029] 5.2) Generate rumor-debunking text: Take the rumor text as input, and calculate and output the rumor-debunking text content of the rumor text through the decoder module of the rumor-debunking text generator; if the input content is not a rumor, directly output the original text, so that the model can output rumor-debunking text for the rumor.

[0030] Compared with the prior art, the advantages of the present invention are:

[0031] The present invention designs a unimodal rumor detection method with automatic rumor refutation capability, which enables the model to complete rumor detection and have automatic rumor refutation capability at the same time. At the same time, the model uses graph learning to extract the causal relationship features of rumor text from the causal graph to enhance the causal semantics of multimodal features, thereby improving the performance of rumor detection.

[0032] While utilizing the causal relationship of rumor text to improve rumor detection performance, the present invention introduces rumor-refuting text generation into the model to enable the model to automatically refute rumors, thereby achieving better performance in single-modal rumor detection and the effect of automatically generating rumor-refuting text. Experiments show that the method performs well in rumor detection tasks, with average classification accuracies of 93.12% and 89.65% on two data sets, respectively, verifying the advantages of causal graphs in improving rumor detection tasks. At the same time, the model can generate rumor-refuting text related to the detection results, improving the interpretability of the detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The present invention is a block diagram of a single-mode rumor detection method with automatic rumor-refuting capability according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the attached embodiment of the present invention Figure 1 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] The embodiment of the present invention comprises the following steps:

[0036] (1) Extract the causal triplets contained in the rumor text in a fine-grained manner and model them in the form of a causal graph;

[0037] (1.1) Through named entity recognition method, named entities in rumor text are extracted, regarded as causal nodes, and causal discovery algorithm is used to mine their causal relationships;

[0038] (1.2) Based on the obtained causal relationship triples, a causal graph related to the rumor is constructed;

[0039] (2) Extracting embedded expression features of causal graphs through graph convolutional neural networks;

[0040] (2.1) Based on the structural information of the causal graph, the adjacency matrix of the graph can be obtained, and the node feature matrix of the graph can be obtained through the word embedding algorithm;

[0041] (2.2) Taking the adjacency matrix and the node feature matrix of the graph as input, the embedded feature representation of the graph can be obtained through the graph convolutional neural network;

[0042] (3) Use the text feature extractor to obtain the features of the rumor text and integrate them with the features of the causal graph to form a causally enhanced single-modal rumor feature;

[0043] (3.1) Using a text feature extractor to obtain feature representation of rumor text;

[0044] (3.2) Fusion of rumor text features with embedded features of causal graphs, and ultimately formation of rumor text features enhanced with causal semantics;

[0045] (4) Using a classifier with an attention mechanism to classify the causally enhanced unimodal features to complete rumor detection;

[0046] (4.1) According to the attention mechanism, the attention weight parameter of the feature is calculated through the feature's Query matrix, Key matrix and Value matrix;

[0047] (4.2) Combined with the attention weight parameter, the rumor classification label obtained after the feature passes through the classifier is obtained, so that the classification result of the rumor text can be obtained.

[0048] (5) Taking the rumor text data as input, passing it through the rumor-refuting text generator substructure in the overall model framework to generate corresponding rumor-refuting text content;

[0049] (5.1) The rumor text is used as the input of the rumor-refuting text generator and passes through its encoder part to form its encoding structure;

[0050] (5.2) The decoder module of the rumor-refuting text generator calculates the rumor-refuting text content of the rumor text.

[0051] If the four indicators of accuracy, precision, recall rate and F1 score are used to evaluate the actual effect of the method of the present invention on single-modal rumor detection, through several experiments, on the English dataset Twitter and the Chinese dataset Weibo, the implementation effect shown in Table 1 is obtained.

[0052] Table 1 Implementation effect (%)

[0053]

[0054] As can be seen from Table 1, the average classification accuracy of this method on the two datasets reached 93.12% and 89.65% respectively. This method can very accurately identify rumor and non-rumor text on the Weibo dataset, with high precision and low missed detection rate; the F1 score of 94.21% is the harmonic mean of precision and recall, which further verifies the excellent performance of the model on the Weibo dataset. Compared with the Weibo dataset, the Twitter dataset of this method still maintains a high level, with accuracy, precision, recall and F1 scores of 89.65%, 88.79%, 89.10% and 88.92% respectively, indicating that this method also has good rumor detection capabilities on the Twitter dataset.

[0055] In this method, the T5 model is used as the rumor-debunking text generator, and the Rouge and Bleu evaluation systems are used to conduct a rumor-debunking text generation quality evaluation experiment, and the implementation results are shown in Table 2.

[0056] Table 2 Effect of rumor-refuting text generation (%)

[0057]

[0058] As can be seen from Table 2, the 94.53% Rouge-1 score indicates that the generated rumor-debunking text has a high degree of overlap with the reference text at the word level, that is, the generated text can well cover the key information of the reference text. The 90.91% Rouge-2 score shows that the overlap at the phrase level is also high, and the generated text is more coherent and accurate in expression. The 94.06% Rouge-L score is based on the longest common subsequence (LCS) evaluation, which further proves the similarity and consistency between the generated rumor-debunking text and the reference text. Although the 87.15% Bleu score is slightly lower than the Rouge index, it is also quite excellent, especially in natural language generation tasks, where Bleu is an important indicator for measuring the n-gram similarity between the generated text and the reference text.

[0059] The following table shows an example of the model’s rumor-refuting text generation.

[0060] Table 3 Examples of rumor-refuting text generation

[0061]

[0062] The above table records the rumor-refuting texts generated by several text generation models for the "injury incident in a certain place". It can be seen from the inference generation results that compared with other models, the T5 of the present invention not only points out the judgment conclusion of the rumor text, but also generates explanatory content for why the incident is judged as a rumor. The generation result is better than ChatGLM and BERT. Therefore, the present invention adopts T5 as the generation model of rumor-refuting text. The present invention enables the model to have the ability to automatically refute rumors while completing rumor detection.

[0063] The above embodiments are only preferred embodiments of the present invention and cannot be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A single-mode rumor detection method with automatic rumor-refuting capability, characterized in that The following steps are involved: 1) Extract the causal triplets contained in the rumor text in a fine-grained manner and model them in the form of a causal graph; 2) Extracting embedded expression features of causal graphs through graph convolutional neural networks; 3) Use the text feature extractor to obtain the features of the rumor text, and integrate them with the features of the causal graph to form a causally enhanced single-modal rumor feature; 4) Using a classifier with an attention mechanism to classify the causally enhanced unimodal features to complete rumor detection; 5) Take the rumor text data as input and pass it through the rumor-refuting text generator substructure in the overall model framework to generate the corresponding rumor-refuting text content.

2. A single-mode rumor detection method with automatic rumor-refuting capability as claimed in claim 1, characterized in that In step 1), the specific method of extracting the causal relationship triples contained in the rumor text in a fine-grained manner and modeling them in the form of a causal graph is: 1.1) Preprocessing: segment and tag rumor text; 1.2) Named Entity Recognition: Named entity recognition technology is used to extract entities in the text, such as names of people, places, and organizations. These entities will serve as potential causal nodes; 1.3) Causal relationship mining: Use causal relationship algorithms to analyze sentences and paragraphs in the text and extract causal relationship triplets; 1.4) Construct a causal graph: Based on the extracted causal triples, a rumor-related causal graph is constructed. The nodes in the graph are entities or events, and the edges are the causal relationships between them.

3. A single-mode rumor detection method with automatic rumor-refuting capability as claimed in claim 1, characterized in that In step 2), the specific method of extracting the embedded expression features of the causal graph through the graph convolutional neural network is: 2.1) Constructing the adjacency matrix: Based on the structural information of the causal graph, obtain its adjacency matrix from the graph ,The elements in the matrix represent the causal relationships between nodes; 2.2) Obtain node feature matrix: For each node, obtain the node set of the graph through the word embedding algorithm The characteristic matrix ; In the formula, Representation Node The Glove algorithm is used to obtain the corresponding word vector features; 2.3) Graph Convolutional Neural Network: Conversion of adjacency matrix and the node feature matrix As input, the graph convolutional neural network (GCN) is used to aggregate information and extract features to obtain the embedded feature representation of each node, that is, in Represents the information aggregation and feature extraction process of graph convolutional neural network.

4. A single-mode rumor detection method with automatic rumor-refuting capability as claimed in claim 1, characterized in that In step 3), the specific method of using a text feature extractor to obtain the features of the rumor text, and fusing them with the features of the causal graph to form a causally enhanced single-modal rumor feature is: 3.1) Text feature extraction: Use the pre-trained text feature extractor Text2Vec to encode the rumor text and obtain the feature vector of the rumor text; 3.2) Feature fusion: The feature vector of the rumor text is fused and spliced ​​with the embedded expression features of the causal graph obtained in step 2) to achieve causal semantic enhancement of the rumor text features and form rumor text features enhanced with causal semantics. , this feature contains both text content information and causal relationship information.

5. A single-mode rumor detection method with automatic rumor refutation capability as claimed in claim 1, characterized in that In step 4), the specific method of using a classifier with an attention mechanism to classify the causally enhanced unimodal features to complete rumor detection is: 4.1) Attention Mechanism: The self-attention mechanism in the Transformer model or a similar attention model is used to calculate the attention weight parameters between features through the query matrix, key matrix, and value matrix of the features to enhance the model's attention to the key semantic content in the text features after causal enhancement; 4.2) Classification: Combined with the attention weight parameter, the attention-weighted features are input into the softmax classifier to complete the classification of rumor text features. The classification is performed and the result of rumor classification is output, that is, whether the text content is a rumor.

6. A single-mode rumor detection method with automatic rumor refutation capability as claimed in claim 1, characterized in that In step 5), the rumor text data is taken as input, and the rumor refuting text generator substructure part in the overall model framework is used to generate the corresponding rumor refuting text content. The specific method is: 5.1) Model selection: According to the characteristics of the rumor-debunking text generation task, select or train a suitable sequence-to-sequence model, such as BERT, T5, or ChatGLM, as a rumor-debunking text generator to generate rumor-debunking text; 5.2) Generate rumor-refuting text: Take the rumor text as input, and calculate and output the rumor-refuting text content of the rumor text through the decoder module of the rumor-refuting text generator; if the input content is not a rumor, the original text is directly output, so that the model can output rumor-refuting text for the rumor.

Images