A text fake information detection method and system fusing contradictory features

Abstract

The application discloses a text false information detection method and system fusing contradictory features, and the method comprises the following steps: performing data preprocessing on a given input text, extracting text features, and extracting sentence pairs with a similarity higher than a threshold to form a similar sentence pair dataset; extracting contradictory word vector features, contradictory scene features and contradictory semantic features in the given input text based on the similar sentence pair dataset; fusing the text features, the contradictory word vector features, the contradictory scene features and the contradictory semantic features, weighting through a self-attention mechanism, and obtaining a weighted and distributed feature fusion vector; and performing false information detection based on the feature fusion vector, and obtaining a false information detection result. The application can fuse contradictory features and style statistical features of the text to perform false information detection, and can effectively improve the accuracy of text-based false information detection.

Description

Technical Field

[0001] This invention relates to the field of text detection technology, and in particular to a method and system for detecting false information in text that integrates contradictory features. Background Technology

[0002] Online social networks (OSNs) have become an important means for people to obtain information, socialize, and share knowledge in their daily lives. In many cases, most ordinary users cannot accurately identify fake news. Manually deleting fake news one by one is time-consuming, laborious, and costly. Therefore, automatically detecting fake news on OSNs has become a current research hotspot.

[0003] Existing methods for detecting misinformation can be divided into two main categories based on whether or not they use external information: methods for detecting misinformation based on the news content itself, and methods for detecting misinformation that incorporate external information.

[0004] Fake news detection methods based on news content can be divided into text-based and multimodal methods. Text-based methods primarily focus on the characteristics of the text content itself, such as text length, style, vocabulary, and statistical or semantic features extracted from the text. Multimodal methods, on the other hand, use a fusion of textual and visual features for classification.

[0005] Fake news detection methods that integrate external information incorporate a wide range of external data. These methods can be categorized into two types based on the external information they rely on: those based on post context and those relying on external knowledge. Context-based methods primarily utilize user feedback, such as user information, comments, likes, shares, and dissemination data, to obtain more comprehensive external features for fake news detection. External knowledge-based methods, on the other hand, rely on existing knowledge bases for fact-checking.

[0006] Existing fake information detection methods based on text content primarily generate a vector representation of each text element before feeding it into a classification layer for true / false classification. Text vector representations can utilize TF-IDF, N-gram, Word2Vec, BERT, and other methods to transform text into vector features. While this text content-based fake information detection has shown good performance, it only utilizes features inherent to the text content itself, such as text length, text style, vocabulary, and statistical or stylistic features extracted from the text content. However, in fake information detection, contradictory features are also a crucial characteristic; fake text often exhibits inconsistencies. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a text misinformation detection method and system that integrates contradictory features. This method can combine contradictory features and stylistic statistical features of text to detect misinformation, thereby improving the detection performance to a certain extent.

[0008] The technical solution adopted in this invention is as follows:

[0009] A text misinformation detection method that integrates contradictory features includes:

[0010] Given input text, perform data preprocessing, extract text features, and extract sentences with similarity higher than a threshold to form a similar sentence pair dataset;

[0011] Based on similar sentences, contradictory word vector features, contradictory scene features, and contradictory semantic features are extracted from the given input text of the dataset.

[0012] Text features, contradictory word vector features, contradictory scene features, and contradictory semantic features are fused together and weighted using a self-attention mechanism to obtain a weighted feature fusion vector.

[0013] False information detection is performed based on feature fusion vectors to obtain false information detection results.

[0014] Furthermore, the extraction of text features includes: mapping the given input text to an input text sequence through a word embedding model, and extracting features from each text vector in the input text sequence through a pre-trained language model, thereby obtaining the text features of the given input text.

[0015] Further, the step of extracting sentences with similarity higher than a threshold to form a similar sentence pair dataset includes: dividing the given input text into multiple sentences to form a sentence set; selecting a statement as a query statement s1, finding a sentence s2 similar to the query statement s1 in the candidate sentence set based on the text matching model, and forming a sentence pair (s1, s2) with the query statement s1; calculating the cosine similarity value of the sentence pair (s1, s2), if it is greater than the threshold, they are similar, otherwise they are not similar; sequentially selecting query statements and calculating the cosine similarity value of the corresponding sentence pairs, and finally forming a similar sentence pair dataset S = {(s1, s2), (s3, s4), …} composed of similar sentence pairs.

[0016] Furthermore, the step of extracting contradictory word vector features from a given input text based on a similar sentence pair dataset includes: training a contradictory word detection model based on the contradictory word dataset to determine whether there is a contradiction between two words; for similar sentences in the similar sentence pair dataset, extracting word pairs that may contain contradictory words and inputting them into the contradictory word detection model to extract contradictory word vector features.

[0017] Furthermore, based on the similar sentence pair dataset, contradictory scene features in the given input text are extracted, including: for similar sentences in the similar sentence pair dataset, scene features corresponding to the sentences are extracted through an event extraction model, and the corresponding scene features of the similar sentence pairs are subtracted to obtain contradictory scene features.

[0018] Furthermore, based on the similar sentence pair dataset, contradictory semantic features are extracted from the given input text. This includes: inputting similar sentences from the similar sentence pair dataset into the semantic feature extraction model, first converting them into vector form through a word embedding layer, then extracting text features through a BiLSTM layer, obtaining interaction information between texts through a collaborative attention layer, obtaining global information of the text through a self-attention layer, and finally obtaining contradictory semantic features through a fusion layer, a pooling layer, and a concatenation layer.

[0019] Furthermore, the process of fusing text features, contradictory word vector features, contradictory scene features, and contradictory semantic features, and then weighting them using a self-attention mechanism, yields a weighted feature fusion vector, including:

[0020] Based on the extracted text features, a text feature vector is obtained through the BERT model. ;

[0021] The contradictory word vector features, contradictory scene features, and contradictory semantic features are concatenated to obtain the contradictory features. ;

[0022] Text feature vectors and contradictory characteristics Feature concatenation is performed to obtain a feature fusion vector. Simultaneously, a weighted feature fusion vector is obtained through selective learning using a self-attention model. .

[0023] Furthermore, the functional expression of the self-attention model includes:

[0024]

[0025] in, For normalized attention weights, query vector Key vector Value vector All are concatenated feature fusion vectors ; This indicates the transpose operation. This is a scaling factor used to stabilize gradient propagation and prevent numerical instability caused by excessively large dot product values. The function is used to convert attention scores into weights, allowing the model to learn the importance of each element pair, thus obtaining the normalized attention weights. ;

[0026] Normalized attention weights Used for feature fusion vectors Perform a weighted summation to generate the final output feature, namely:

[0027]

[0028] in, This is the weighted feature fusion vector.

[0029] Furthermore, the step of detecting false information based on feature fusion vectors and obtaining false information detection results includes:

[0030] Based on The activated fully connected layer fuses the feature vector. Projecting onto two target spaces—real information and false information—simultaneously yields probability distributions:

[0031]

[0032] in, It is a prediction probability vector. It is the probability that the prediction is true. It is the probability that the prediction is false information; It is a weight matrix. It is a bias term;

[0033] The training process aims to minimize the cross-entropy loss function.

[0034]

[0035] in, For the purpose of loss, For all training parameters of the neural network, For predicting labels, For real labels, For the number of training sessions.

[0036] A text misinformation detection system that integrates contradictory features includes:

[0037] The data preprocessing module is configured to preprocess the given input text.

[0038] The text feature extraction module is configured to extract text features from a given input text;

[0039] The sentence similarity extraction module is configured to extract sentences with a similarity higher than a threshold from a given input text and form a dataset of similar sentence pairs.

[0040] The contradiction feature extraction module is configured to extract contradictory word vector features, contradictory scene features, and contradictory semantic features from a given input text based on a dataset of similar sentence pairs.

[0041] The feature fusion module is configured to fuse text features, contradictory word vector features, contradictory scene features, and contradictory semantic features, and then weight them through a self-attention mechanism to obtain a weighted feature fusion vector.

[0042] The fake information detection module is configured to detect fake information based on feature fusion vectors and obtain the fake information detection results.

[0043] The beneficial effects of this invention are as follows:

[0044] To address the phenomenon of conflicting false and true information, this invention integrates contradictory features of text content, extracting contradictory word vector features, contradictory scene features, contradictory semantic features, as well as statistical and stylistic features of the text content itself. Simultaneously, it utilizes an attention mechanism to effectively fuse contradictory features and text features to complete false information detection, thereby significantly improving the accuracy of text-based false information detection. Attached Figure Description

[0045] Figure 1 This is a flowchart of a text false information detection method that integrates contradictory features according to Embodiment 1 of the present invention.

[0046] Figure 2 This is a schematic diagram of the contradiction word detection model in Embodiment 1 of the present invention.

[0047] Figure 3 This is a schematic diagram of the semantic feature extraction model of Embodiment 1 of the present invention.

[0048] Figure 4 This is a schematic diagram of a text misinformation detection system that integrates contradictory features, according to Embodiment 2 of the present invention. Detailed Implementation

[0049] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] like Figure 1As shown, this embodiment provides a text misinformation detection method that integrates contradictory features, including:

[0052] Given input text, perform data preprocessing, extract text features, and extract sentences with similarity higher than a threshold to form a similar sentence pair dataset;

[0053] Based on similar sentences, contradictory word vector features, contradictory scene features, and contradictory semantic features are extracted from the given input text of the dataset.

[0054] Text features, contradictory word vector features, contradictory scene features, and contradictory semantic features are fused together and weighted using a self-attention mechanism to obtain a weighted feature fusion vector.

[0055] False information detection is performed based on feature fusion vectors to obtain false information detection results.

[0056] Preferably, in the data preprocessing stage, the dataset in this embodiment uses publicly available fake information datasets (21 Chinese Weibo posts, 15 / 16 English Twitter posts) and labeled datasets crawled from Snopes, PolitiFact, and Tencent's "Fact Check" debunking mini-program. For text data, due to its complex information and high redundancy, data cleaning is performed for consistency. Specifically, emoticons, URLs, and special symbols (such as '###') are removed from the text; spaces and whitespace characters are deleted and replaced; if the text contains traditional Chinese characters, they are first converted to simplified Chinese, resulting in a cleaned text dataset.

[0057] Preferably, in the text feature extraction stage, this embodiment maps the given input text into an input text sequence through a word embedding model, and extracts features from each text vector in the input text sequence through a pre-trained language model (e.g., the BERT model), thereby obtaining the text features of the given input text.

[0058] Since sentences with high similarity may contain contradictions, it is necessary to first extract sentences with high similarity to form a similar sentence pair dataset. Preferably, the given input text is divided into multiple sentences to form a sentence set; a statement is selected as the query statement s1, and based on the text matching model, a sentence s2 similar to the query statement s1 is found in the candidate sentence set, and a sentence pair (s1, s2) is formed with the query statement s1; the cosine similarity value of the sentence pair (s1, s2) is calculated, and if it is greater than a threshold (e.g., 0.7), they are considered similar, otherwise they are not similar; query statements are selected in turn, and the cosine similarity value of the corresponding sentence pairs is calculated, finally forming a similar sentence pair dataset S = {(s1, s2), (s3, s4), …} composed of similar sentence pairs.

[0059] Preferably, in the contradictory word vector feature extraction stage, a contradictory word detection model is trained based on the contradictory word dataset to determine whether there is a contradiction between two words; for similar sentences in the similar sentence pair dataset, word pairs that may contain contradictory words are extracted and input into the contradictory word detection model to extract contradictory word vector features.

[0060] Specifically, after obtaining the input text to be detected, the sentence similarity extraction module first extracts potentially contradictory similar sentence pairs. Then, based on these similar sentence pairs, difflib is used to obtain the difference words between the sentence pairs. These difference words are used as the input word pair for the contradiction word detection model, which then extracts the contradiction features of the word pair. The contradiction word detection model is a Siamese network structure, with the specific structure as follows: Figure 2 As shown, word pairs in the text are used to extract features using a shared BERT model. Then, the two features are subtracted to obtain contradictory word vector features.

[0061] Preferably, in the contradiction scene feature extraction stage, for similar sentences in the similar sentence pair dataset, the scene features corresponding to the sentences are extracted by the event extraction model, and the corresponding scene features of the similar sentence pairs are subtracted to obtain the contradiction scene features.

[0062] Specifically, after obtaining the input text to be detected, the sentence similarity extraction module first extracts potentially contradictory similar sentence pairs; then, based on these similar sentence pairs, scene feature extraction is performed. The scene feature extraction model selects the open-source event information extraction toolkit OmniEvent, using the text of the text pairs as the model input, selecting 'EAE' (event extraction mode) as the task mode, and extracting the 'time' and 'place' attribute features as the extracted scene features. The corresponding scene features of the similar sentence pairs are subtracted to obtain contradictory scene features.

[0063] Preferably, in the stage of extracting contradictory semantic features, similar sentences in the similar sentence pair dataset are input into the semantic feature extraction model. They are first converted into vector form through a word embedding layer, then the text features are extracted through a BiLSTM layer, the interaction information between texts is obtained through a collaborative attention layer, and the global information of the text is obtained through a self-attention layer. Finally, contradictory semantic features are obtained through a fusion layer, a pooling layer, and a concatenation layer.

[0064] Specifically, after obtaining the input text to be detected, the sentence similarity extraction module first extracts potentially contradictory similar sentence pairs; then, based on these similar sentence pairs, semantic feature extraction is performed. The semantic feature extraction model is as follows: Figure 3As shown, the input text is first converted into vector form through a word embedding layer, then the text features are extracted through a BiLSTM layer, the interaction information between texts is obtained through a collaborative attention layer, the global information of the text is obtained through a self-attention layer, and finally the contradictory semantic features are obtained by concatenating the fusion layer and the pooling layer.

[0065] Preferably, the feature fusion stage can be implemented using the following steps:

[0066] Based on the extracted text features, a text feature vector is obtained through the BERT model. ;

[0067] The contradictory word vector features, contradictory scene features, and contradictory semantic features are concatenated to obtain the contradictory features. ;

[0068] Text feature vectors and contradictory characteristics Feature concatenation is performed to obtain a feature fusion vector. Simultaneously, a weighted feature fusion vector is obtained through selective learning using a self-attention model. .

[0069] It should be noted that the functional expression of the self-attention model is:

[0070]

[0071] in, For normalized attention weights, query vector Key vector Value vector All are concatenated feature fusion vectors ; This indicates the transpose operation. This is a scaling factor used to stabilize gradient propagation and prevent numerical instability caused by excessively large dot product values. The function is used to convert attention scores into weights, allowing the model to learn the importance of each element pair, thus obtaining the normalized attention weights. ;

[0072] Normalized attention weights Used for feature fusion vectors Perform a weighted summation to generate the final output feature, namely:

[0073]

[0074] in, This is the weighted feature fusion vector.

[0075] Preferably, the false information stage can be implemented using the following steps:

[0076] Based on The activated fully connected layer fuses the feature vector. Projecting onto two target spaces—real information and false information—simultaneously yields probability distributions:

[0077]

[0078] in, It is a prediction probability vector. It is the probability that the prediction is true. It is the probability that the prediction is false information; It is a weight matrix. It is a bias term;

[0079] The training process aims to minimize the cross-entropy loss function.

[0080]

[0081] in, For the purpose of loss, For all training parameters of the neural network, For predicting labels, For real labels, For the number of training sessions.

[0082] Example 2

[0083] like Figure 4 As shown, this embodiment provides a text misinformation detection system that integrates contradictory features, including:

[0084] The data preprocessing module is configured to preprocess the given input text.

[0085] The text feature extraction module is configured to extract text features from a given input text;

[0086] The sentence similarity extraction module is configured to extract sentences with a similarity higher than a threshold from a given input text and form a dataset of similar sentence pairs.

[0087] The contradictory word vector feature extraction module is configured to extract contradictory word vector features from a given input text based on a similar sentence pair dataset.

[0088] The contradictory scene feature extraction module is configured to extract contradictory scene features from a given input text based on a dataset of similar sentence pairs.

[0089] The contradictory semantic feature extraction module is configured to extract contradictory semantic features from a given input text based on a dataset of similar sentence pairs.

[0090] The feature fusion module is configured to fuse text features, contradictory word vector features, contradictory scene features, and contradictory semantic features, and then weight them through a self-attention mechanism to obtain a weighted feature fusion vector.

[0091] The fake information detection module is configured to detect fake information based on feature fusion vectors and obtain the fake information detection results.

[0092] Example 3

[0093] This embodiment is based on embodiment 1:

[0094] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the text misinformation detection method incorporating contradictory features as described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0095] Example 4

[0096] This embodiment is based on embodiment 1:

[0097] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the text misinformation detection method incorporating contradictory features described in Embodiment 1. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content contained in the storage medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the storage medium does not include electrical carrier signals and telecommunication signals.

[0098] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A method for detecting textual misinformation that integrates contradictory features, characterized in that, include: Given input text, perform data preprocessing, extract text features, and extract sentences with similarity higher than a threshold to form a similar sentence pair dataset; Based on similar sentences, contradictory word vector features, contradictory scene features, and contradictory semantic features are extracted from the given input text of the dataset. Text features, contradictory word vector features, contradictory scene features, and contradictory semantic features are fused together and weighted using a self-attention mechanism to obtain a weighted feature fusion vector. False information detection is performed based on feature fusion vectors to obtain false information detection results; The extraction of text features includes: mapping a given input text to an input text sequence using a word embedding model, and extracting features from each text vector in the input text sequence using a pre-trained language model, thereby obtaining the text features of the given input text.

2. The textual false information detection method integrating contradictory features according to claim 1, characterized in that, The extracted sentences with similarity higher than a threshold are used to form a similar sentence pair dataset, including: Divide the given input text into multiple sentences to form a sentence set; Select a statement as query statement s1. Based on the text matching model, find a sentence s2 that is similar to query statement s1 in the candidate sentence set and form a sentence pair (s1, s2) with query statement s1. Calculate the cosine similarity value of sentence pair (s1, s2). If it is greater than the threshold, they are similar; otherwise, they are not similar. The query statements are selected sequentially, and the cosine similarity values ​​of the corresponding sentence pairs are calculated. Finally, a similar sentence pair dataset S = {(s1, s2), (s3, s4), …} is constructed.

3. The textual false information detection method integrating contradictory features according to claim 1, characterized in that, The step of extracting contradictory word vector features from a given input text based on a similar sentence pair dataset includes: training a contradictory word detection model based on the contradictory word dataset to determine whether there is a contradiction between two words; and for similar sentences in the similar sentence pair dataset, extracting word pairs that may contain contradictory words and inputting them into the contradictory word detection model to extract contradictory word vector features.

4. The textual false information detection method integrating contradictory features according to claim 1, characterized in that, Extracting contradictory scene features from a given input text based on a similar sentence pair dataset includes: extracting scene features corresponding to similar sentences in the similar sentence pair dataset through an event extraction model, and subtracting the corresponding scene features of the similar sentence pairs to obtain contradictory scene features.

5. The textual false information detection method integrating contradictory features according to claim 1, characterized in that, The method for extracting contradictory semantic features from a given input text based on a similar sentence pair dataset includes: inputting similar sentences from the similar sentence pair dataset into the semantic feature extraction model, first converting them into vector form through a word embedding layer, then extracting text features through a BiLSTM layer, obtaining interaction information between texts through a collaborative attention layer, obtaining global information of the text through a self-attention layer, and finally obtaining contradictory semantic features through a fusion layer, a pooling layer, and a concatenation layer.

6. The textual false information detection method integrating contradictory features according to claim 1, characterized in that, The process involves fusing text features, contradictory word vector features, contradictory scene features, and contradictory semantic features, and then weighting them using a self-attention mechanism to obtain a weighted feature fusion vector, including: Based on the extracted text features, the text feature vector is obtained through the BERT model. ; The contradictory word vector features, contradictory scene features, and contradictory semantic features are concatenated to obtain the contradictory features. ; Text feature vectors and contradictory characteristics Feature concatenation is performed to obtain a feature fusion vector. Simultaneously, a weighted feature fusion vector is obtained through selective learning using a self-attention model. .

7. The textual false information detection method integrating contradictory features according to claim 6, characterized in that, The functional expression of the self-attention model includes: in, For normalized attention weights, query vector Key vector Value vector All are concatenated feature fusion vectors ; This indicates the transpose operation. This is a scaling factor used to stabilize gradient propagation and prevent numerical instability caused by excessively large dot product values. The function is used to convert attention scores into weights, allowing the model to learn the importance of each element pair, thus obtaining the normalized attention weights. ; Normalized attention weights Used for feature fusion vectors Perform a weighted summation to generate the final output feature, namely: in, This is the weighted feature fusion vector.

8. The textual false information detection method integrating contradictory features according to claim 1, characterized in that, The method of detecting false information based on feature fusion vectors and obtaining false information detection results includes: Based on The activated fully connected layer fuses the feature vector. Projecting onto two target spaces—real information and false information—simultaneously yields probability distributions: in, It is a prediction probability vector. It is the probability that the prediction is true. It is the probability that the prediction is false information; It is a weight matrix. It is a bias term; The training process aims to minimize the cross-entropy loss function. in, For the purpose of loss, For all training parameters of the neural network, For predicting labels, For real labels, This represents the number of samples.

9. A text-based false information detection system that integrates contradictory features, characterized in that, include: The data preprocessing module is configured to preprocess the given input text. The text feature extraction module is configured to extract text features from a given input text; The sentence similarity extraction module is configured to extract sentences with a similarity higher than a threshold from a given input text and form a dataset of similar sentence pairs. The contradiction feature extraction module is configured to extract contradictory word vector features, contradictory scene features, and contradictory semantic features from a given input text based on a dataset of similar sentence pairs. The feature fusion module is configured to fuse text features, contradictory word vector features, contradictory scene features, and contradictory semantic features, and then weight them through a self-attention mechanism to obtain a weighted feature fusion vector. The fake information detection module is configured to detect fake information based on feature fusion vectors and obtain the fake information detection results. The extraction of text features from a given input text includes: mapping the given input text to an input text sequence using a word embedding model, and extracting features from each text vector in the input text sequence using a pre-trained language model, thereby obtaining the text features of the given input text.

Images