Rumor detection method based on topic audio frequency

Abstract

The invention belongs to the technical field of natural language processing and information processing, and particularly relates to a topic audio-based rumor detection method, which comprises the following steps of: processing acquired social propagation data to obtain a comment sequence; processing the comment sequence by using a pre-trained word embedding model to obtain an embedded representation sequence, and processing the embedded representation sequence by using an emotion prediction model to obtain an emotion feature sequence; constructing a similar audio signal fragment sequence according to the emotion feature sequence; splicing and converting all class audio signal fragments in the class audio signal fragment sequence to obtain discrete class audio signals; in combination with Fourier transform and a Mel spectrum algorithm, acquiring similar audio features of the discrete similar audio signals; inputting the audio-like features into a prediction model to obtain a prediction result; calculating loss according to a prediction result and carrying out back propagation training until the training is completed; according to the method, a lightweight rumor detection model is established, cross-domain complex and changeable social rumor tasks can be dealt with, and rumor propagation can be inhibited.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing and information processing technology, and specifically relates to a rumor detection method based on topic audio. Background Art

[0002] With the rapid development of the internet, the number of netizens has increased dramatically, making their identities increasingly difficult to determine. Numerous online influencer accounts, known as "troll army" accounts, have been spreading rumors to manipulate public opinion, leading to major public opinion crises. The industry has implemented numerous measures to detect rumors and mitigate these crises. For example, social media platforms such as Weibo and TikTok have launched rumor-refuting platforms. Traditional methods extract artificial features from topic content and user information, then use these features to train detection models based on decision trees or support vector machines. With the advancement of deep learning, a growing number of deep learning-based rumor detection methods have emerged, primarily falling into two categories. On the one hand, neural network models are used to extract semantic, word frequency, and signature word features from topic content to detect rumors. Furthermore, with the advancement of computer vision (CV), methods incorporating multimedia features are also increasing. On the other hand, graph neural network models (GNNs) are used to extract more advanced representations of propagation paths and network structures from transmitted information.

[0003] Among existing algorithms, those based on machine learning and basic deep learning lack performance capabilities and are therefore unable to meet current industry demands. More advanced algorithms (CV and GNN) meet performance requirements, but their models are large when faced with massive amounts of user data, and online deployment consumes significant computing power. Therefore, a new rumor detection approach is urgently needed, one that can build a lightweight model while maintaining high detection performance. Summary of the Invention

[0004] To solve the above problems, the present invention provides a rumor detection method based on topic audio, comprising the following steps:

[0005] S1. Process the collected social communication data to obtain a comment sequence;

[0006] S2. Use the pre-trained word embedding model to process the comment sequence to obtain an embedded representation sequence, and then use the sentiment prediction model to process the embedded representation sequence to obtain a sentiment feature sequence;

[0007] S3. Constructing a sequence of audio signal segments based on the emotional feature sequence;

[0008] S4. All audio signal segments in the audio signal segment sequence are concatenated and converted to obtain a discrete audio signal;

[0009] S5. Combining Fourier transform and Mel spectrum algorithm to obtain audio-like features of discrete audio signals;

[0010] S6. Inputting the audio-like features into a prediction model to obtain a prediction result, wherein the prediction model is constructed using a convolutional neural network;

[0011] S7. Calculate the loss based on the prediction results and perform backpropagation training. Repeat steps S1-S7 until the training is completed.

[0012] Beneficial effects of the present invention:

[0013] The present invention first converts the topic space into an audio-like signal. This method quantifies the characteristics of topic propagation and maps them to the amplitude and frequency characteristics of the audio-like signal. Secondly, Fourier transform and Mel-spectrogram algorithms are used to retrieve audio features. Finally, a convolutional neural network is used to deeply mine the audio-like features, followed by two linear layers to predict topic tags. This invention establishes a lightweight rumor detection model that can handle complex and diverse social rumor tasks across multiple domains, helping to curb the spread of rumors. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flowchart for training the enhanced multi-relation rumor detection model in the present invention. DETAILED DESCRIPTION

[0015] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0016] The present invention provides a rumor detection method based on topic audio. Figure 1 As shown, the following steps are included:

[0017] S1. Process the collected social communication data to obtain a comment sequence.

[0018] Specifically, the collected social communication data includes title tags (whether it is a rumor), topic text, all user comment texts in the topic discussion thread, release time, number of user followers, and whether the user is an official account.

[0019] Sort all user comment texts by release time to obtain a comment sequence.

[0020] S2. Use the pre-trained word embedding model to process the comment sequence to obtain the embedded representation sequence, and then use the sentiment prediction model to process the embedded representation sequence to obtain the sentiment feature sequence.

[0021] Specifically, the pre-trained word embedding model in step S2 adopts the Bert model, and the processing process is expressed as follows:

[0022]

[0023] Among them, C represents the comment sequence, Representation embedding represents sequence; Indicates the length of the comment sequence, represents the embedding dimension of a single review, and .

[0024] The sentiment prediction model uses a long short-term memory neural network to obtain the sentiment characteristics of each comment in the comment sequence. The processing process is expressed as

[0025]

[0026] in, represents a sequence of sentiment features; LSTM( ) represents a long short-term memory neural network, whose output layer outputs the predicted feature values ​​of three categories of sentiment: positive, negative, and neutral. softmax( ) represents the softmax function.

[0027] S3. Construct a sequence of audio signal segments based on the emotional feature sequence.

[0028] Specifically, step S3 includes:

[0029] S31. Calculate the frequency characteristics of the audio signal segment of each comment in the comment sequence based on the emotional feature sequence, expressed as

[0030]

[0031] Among them, E in F represents the nth type of emotion of the i-th element in the emotion feature sequence, that is, the nth type of emotion of the emotion feature of the i-th comment in the comment sequence. Specifically, the emotion feature of each comment has three types of emotion characteristics, n=1 represents positive emotion, n=2 represents neutral emotion, and n=3 represents negative emotion; i The frequency characteristics of the audio-like signal segment representing the i-th comment in the comment sequence;

[0032] S32. Calculate the amplitude feature of the audio signal segment of each comment in the comment sequence, expressed as

[0033]

[0034] Among them, A i The amplitude feature of the audio signal segment representing the i-th comment in the comment sequence; useroff iIndicates whether the user account corresponding to the i-th comment in the comment sequence is an official account. If so, useroff i =1, if not, then useroff i =0;fansNum i represents the number of fans of the user account corresponding to the i-th comment in the comment sequence, fansThre represents the threshold, and φ represents the attenuation factor;

[0035] S33. Construct a similar audio signal segment according to the frequency characteristics of the similar audio signal segment and the amplitude characteristics of the similar audio signal segment, expressed as

[0036]

[0037] Among them, Audio i (t) represents the audio signal segment corresponding to the i-th comment, t is the time variable, and cos represents the cosine function.

[0038] S4. Splicing and converting all the similar audio signal segments in the similar audio signal segment sequence to obtain discrete similar audio signals.

[0039] Specifically, step S4 includes:

[0040] S41. Adjust all audio signal segments to the same time length;

[0041] In order to ensure the consistency of the length of the audio signal, the time length of the complete audio signal is defined as , and then dynamically calculate the time length of each audio signal segment, expressed as:

[0042]

[0043] in, Indicates the time length of a single audio signal segment. Indicates the time length of the complete audio signal. Indicates the length of the comment sequence.

[0044] S42. Splice all similar audio signal segments into a complete similar audio signal Audio ( t )= Concat( { Audio 1 ( t ), ..., Audio s ( t )}) t∈[0, T]

[0045] Where t represents the time variable, T represents the length of the complete audio-like signal, Concat represents the concatenation function, and s represents the number of audio-like signal segments;

[0046] S43. Convert continuous complete audio signals into discrete audio signals according to the sampling theorem

[0047] in, Indicates the interval is The sampling function of Represents a convolution operation.

[0048] S5. Combine Fourier transform and Mel spectrum algorithm to obtain audio-like features of discrete audio-like signals.

[0049] Specifically, step S5 includes:

[0050] S51. Use fast Fourier transform to convert discrete audio signals into frequency domain;

[0051] S52. Use the Mel spectrum algorithm to process the discrete audio-like signal converted into the frequency domain to obtain audio-like features.

[0052] The above process can be expressed as

[0053]

[0054] in, Represents audio-like features. Represents a discrete class audio signal. and Represents the Fast Fourier Transform and Mel Spectrum algorithms.

[0055] S6. Input the audio-like features into a prediction model to obtain a prediction result, where the prediction model is constructed using a convolutional neural network.

[0056] Specifically, step S6 inputs the audio-like features into the prediction model to obtain the prediction result expressed as

[0057]

[0058]

[0059]

[0060] Where P represents the prediction result, H represents the audio-like feature, CNN( ) represents the convolutional neural network, Flatten( ) represents the one-dimensional expansion algorithm, ReLu( ) represents the activation function, softmax( ) represents the activation function, Dropout( ) represents the Dropout algorithm, W1 and W2 represent weight matrices, and b1 and b2 represent biases. represents the output of the convolutional neural network, Represents the output vector of the first linear layer. And in order to prevent overfitting during training, the present invention uses algorithm, further enhancing the generalization of the model.

[0061] S7. Calculate the loss based on the prediction results and perform backpropagation training. Repeat steps S1-S7 until the training is completed.

[0062] Specifically, the model loss function uses cross entropy loss:

[0063]

[0064] in, represents the total loss of the model, represents the true label of the i-th topic, Represents the rumor prediction result of the i-th topic.

[0065] Based on the model loss, the neural network backpropagation mechanism is used to train the prediction model and update the parameters. After the loss function converges or reaches the predetermined number of iterations, the trained prediction model is obtained. The word embedding model, sentiment prediction model, feature extraction conversion module (steps S3-S5) and prediction model are combined to obtain an enhanced multi-relation rumor detection model, which can accurately identify whether a topic is a rumor.

[0066] By obtaining the social communication data to be identified and inputting it into the trained enhanced multi-relation rumor detection model for processing, the rumor detection results of the topics in the social communication data can be obtained.

[0067] In the present invention, unless otherwise clearly stipulated and limited, the terms "installation", "setting", "connection", "fixation", "rotation" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0068] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A rumor detection method based on topic audio, characterized in that: The following steps are involved: S1. Process the collected social communication data to obtain a comment sequence; S2. Use the pre-trained word embedding model to process the comment sequence to obtain an embedded representation sequence, and then use the sentiment prediction model to process the embedded representation sequence to obtain a sentiment feature sequence; S3. Constructing a sequence of audio signal segments based on the emotional feature sequence; S4. All audio signal segments in the audio signal segment sequence are concatenated and converted to obtain a discrete audio signal; S5. Combining Fourier transform and Mel spectrum algorithm to obtain audio-like features of discrete audio signals; S6. Inputting the audio-like features into a prediction model to obtain a prediction result, wherein the prediction model is constructed using a convolutional neural network; S7. Calculate the loss based on the prediction results and perform backpropagation training. Repeat steps S1-S7 until the training is completed.

2. The rumor detection method based on topic audio according to claim 1 is characterized in that: The pre-trained word embedding model in step S2 adopts the Bert model, and the processing process is expressed as Among them, C represents the review sequence, R represents the embedding representation sequence; The sentiment prediction model uses a long short-term memory neural network to obtain the sentiment characteristics of each comment in the comment sequence. The processing process is expressed as Where E represents the emotional feature sequence, LSTM( ) represents the long short-term memory neural network, and softmax( ) represents the softmax function.

3. The rumor detection method based on topic audio according to claim 1 is characterized in that: Step S3 specifically includes: S31. Calculate the frequency characteristics of the audio signal segment of each comment in the comment sequence based on the emotional feature sequence, expressed as Among them, E in F represents the nth type of emotion of the i-th element in the emotion feature sequence, that is, the nth type of emotion of the emotion feature of the i-th comment in the comment sequence. When n=1, it represents positive emotion, when n=2, it represents neutral emotion, and when n=3, it represents negative emotion. i The frequency characteristics of the audio-like signal segment representing the i-th comment in the comment sequence; S32. Calculate the amplitude feature of the audio signal segment of each comment in the comment sequence, expressed as Among them, A i The amplitude feature of the audio signal segment representing the i-th comment in the comment sequence; useroff i Indicates whether the user account corresponding to the i-th comment in the comment sequence is an official account. If so, useroff i =1, if not, then useroff i =0;fansNum i represents the number of fans of the user account corresponding to the i-th comment in the comment sequence, fansThre represents the threshold, and φ represents the attenuation factor; S33. Construct a similar audio signal segment according to the frequency characteristics of the similar audio signal segment and the amplitude characteristics of the similar audio signal segment, expressed as Among them, Audio i (t) represents the audio signal segment corresponding to the i-th comment, t is the time variable, and cos represents the cosine function.

4. The rumor detection method based on topic audio according to claim 1 is characterized in that: Step S4 includes: S41. Adjust all audio signal segments to the same time length; S42. Splice all similar audio signal segments into a complete similar audio signal Where t represents the time variable, T represents the length of the complete audio-like signal, Concat represents the concatenation function, and s represents the number of audio-like signal segments; S43. Convert continuous complete audio signals into discrete audio signals according to the sampling theorem in, Indicates the interval is The sampling function of Represents a convolution operation.

5. The rumor detection method based on topic audio according to claim 1 is characterized in that: Step S5 includes: S51. Use fast Fourier transform to convert discrete audio signals into frequency domain; S52. Use the Mel spectrum algorithm to process the discrete audio-like signal converted into the frequency domain to obtain audio-like features.

6. The rumor detection method based on topic audio according to claim 1 is characterized in that: Step S6 inputs the audio-like features into the prediction model to obtain the prediction result expressed as Where P represents the prediction result, H represents the audio-like feature, CNN( ) represents the convolutional neural network, Flatten( ) represents the one-dimensional expansion algorithm, ReLu( ) represents the activation function, softmax( ) represents the activation function, Dropout( ) represents the Dropout algorithm, W1 and W2 represent weight matrices, and b1 and b2 represent biases.

Images