Network spoofing detection system and method based on deep learning and value decision

Abstract

The invention discloses a network spoofing detection system and method based on deep learning and value decision, and relates to the technical field of natural language processing, and the method comprises the steps: obtaining to-be-detected original text data, and carrying out the preprocessing, and obtaining a text identifier sequence; inputting the text identifier sequence into a fine-tuned pre-training language model, and extracting a global semantic feature vector of the original text data; inputting the global semantic feature vector into a multi-channel convolutional neural network to obtain a fused feature vector; inputting the fusion feature vector into a trained actor and commentator network, and outputting a prediction probability of network spoofing; and according to the prediction probability, determining whether the original text data is network spoofing. According to the method, the global and local semantic features are fused, and the actor-commentator network is introduced to realize dynamic optimization, so that the problems of single feature utilization and insufficient complex text recognition capability of the existing method are solved, and the detection precision and the adaptability to fuzzy spoofing expression are improved.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a cyberbullying detection system and method based on deep learning and value decision-making. Background Technology

[0002] With the widespread adoption of the internet and social media, user-generated content has exploded. While social networks have become important sources of information and communication tools, they have also exacerbated the spread of cyberbullying. Cyberbullying is characterized by its lack of time and space constraints, anonymity, rapid spread, and public persistence. Its harm has affected multiple groups, including teenagers, young adults, and working professionals, and has become a globally significant social problem. To address cyberbullying in a timely manner, automated cyberbullying detection technology, due to its efficiency and real-time capabilities, has become a research hotspot in academia and industry.

[0003] Early research relied primarily on traditional machine learning algorithms, such as Support Vector Machines (SVM), Naive Bayes (NB), and Decision Trees (DT). These algorithms required manual design of features like word frequency and TF-IDF, making it difficult to capture complex semantics in text and resulting in limited detection effectiveness. Subsequently, deep learning methods were introduced into the field, such as Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). However, these models have fixed parameters after training, making it difficult to effectively identify ambiguous expressions commonly found in cyberbullying, such as sarcasm, irony, and veiled threats, which can easily lead to misjudgments.

[0004] Existing technologies fail to effectively integrate the global contextual semantics of text with multi-scale local information, resulting in insufficient ability to identify complex cyberbullying texts. Summary of the Invention

[0005] The purpose of this application is to provide a cyberbullying detection system and method based on deep learning and value decision-making. By integrating global and local semantic features and introducing an actor-critic network to achieve dynamic optimization, it solves the problems of existing methods that rely on single features and rigid decision-making, resulting in insufficient recognition ability for complex texts. This significantly improves the detection accuracy and adaptability to ambiguous bullying expressions.

[0006] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for detecting cyberbullying based on deep learning and value decision-making, comprising: acquiring raw text data to be detected; preprocessing the raw text data to obtain a sequence of text identifiers; inputting the text identifier sequence into a fine-tuned pre-trained language model, and using the fine-tuned pre-trained language model to extract a global semantic feature vector of the raw text data; inputting the global semantic feature vector into a multi-channel convolutional neural network, and extracting local features of different scales through multiple parallel convolutional layers in the multi-channel convolutional neural network, and fusing the local features of different scales to obtain a fused feature vector; inputting the fused feature vector into a trained actor critic network, and outputting a predicted probability that the raw text data constitutes cyberbullying through the actor network in the trained actor critic network; and determining whether the raw text data constitutes cyberbullying based on the predicted probability.

[0007] Optionally, the training process of the trained actor-critic network includes: simultaneously inputting the fused feature vector of the training sample into the unoptimized actor-critic network; outputting the predicted probability of the training sample through the actor network; and outputting the state evaluation value of the training sample in the current predicted state through the critic network; using the true label of the training sample as the reward signal and the state value evaluation to calculate the advantage value; and optimizing the parameters of the unoptimized actor network and the critic network based on the advantage value and the predicted probability of the training sample, respectively.

[0008] Optionally, optimizing the parameters of the unoptimized actor network and critic network based on the advantage value and the predicted probability of the training samples includes: constructing a loss function for the actor network based on the advantage value and the predicted probability of the training samples; constructing a loss function for the critic network based on the advantage value; constructing a total loss function by weighted sum of the loss functions of the actor network and the critic network; and updating the parameters of the actor network and the critic network using the total loss function.

[0009] Optionally, the formula for constructing the loss function of the actor network is L. actor =-log(π(a|s))×A(s,a), where, L actor Let π(a|s) be the loss function of the actor network, π(a|s) be the predicted probability of the sample, and A(s,a) be the advantage value.

[0010] Optionally, the loss function of the critic network is constructed using the formula L. critic =[A(s,a)] 2 , where L critic Let A(s,a) be the loss function of the critic network, and let A(s,a) be the advantage value.

[0011] Optionally, the advantage value is calculated using the formula A(s,a)=R(s,a)-V(s), where A(s,a) is the advantage value, R(s,a) is the reward signal, and V(s) is the state value assessment.

[0012] Optionally, the multichannel convolutional neural network includes three parallel convolutional layers with different kernel sizes, which are used to extract local features at different scales respectively.

[0013] Optionally, the pre-trained language model is a RoBERTa pre-trained language model.

[0014] Secondly, this application provides a cyberbullying detection system based on deep learning and value decision-making, comprising: an acquisition module configured to acquire raw text data to be detected; a preprocessing module configured to preprocess the raw text data to obtain a sequence of text identifiers; a feature vector extraction module configured to input the text identifier sequence into a fine-tuned pre-trained language model and extract global semantic feature vectors of the raw text data using the fine-tuned pre-trained language model; a feature fusion module configured to input the global semantic feature vectors into a multi-channel convolutional neural network, extract local features of different scales through multiple parallel convolutional layers in the multi-channel convolutional neural network, and fuse the local features of different scales to obtain a fused feature vector; a prediction probability generation module configured to input the fused feature vectors into a trained actor critic network and output a prediction probability that the raw text data constitutes cyberbullying through the actor network in the trained actor critic network; and a bullying determination module configured to determine whether the raw text data constitutes cyberbullying based on the prediction probability.

[0015] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the cyberbullying detection method based on deep learning and value decision-making as described above.

[0016] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a cyberbullying detection system and method based on deep learning and value decision-making. By preprocessing the original text data, it addresses the problem of noise in original social media text data, achieving standardization of the input data. Through global semantic feature extraction based on a pre-trained language model, it overcomes the limitations of existing cyberbullying detection methods in deeply understanding textual context and capturing long-distance dependencies, enabling the capture of global deep semantic information such as insulting intent and aggressive context, thus improving semantic understanding capabilities. Furthermore, by using multi-scale local feature fusion based on a multi-channel convolutional neural network, it addresses the insufficient ability of a single pre-trained model to capture key local information, thus improving the ability to capture local key information. Key information such as specific abusive words and phrase combinations are extracted, and the global semantic feature extraction of the pre-trained language model is combined with the multi-scale local feature fusion of the multi-channel convolutional neural network. This solves the problem of single feature extraction in existing methods, and achieves complementary fusion of global contextual semantics and multi-scale local key patterns, enhancing the model's ability to distinguish complex bullying expressions. By using a trained actor network, which has internalized a dynamic optimization strategy based on value assessment, it can output a predicted probability optimized by value decision, ultimately improving the overall detection performance. This enables more accurate identification of cyberbullying behavior and significantly reduces the false positive rate, especially improving adaptability to ambiguous language scenarios and enhancing generalization ability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an application environment diagram of a cyberbullying detection method based on deep learning and value decision-making in one embodiment of this application; Figure 2 A flowchart illustrating a cyberbullying detection method based on deep learning and value decision-making, provided as an embodiment of this application; Figure 3 A flowchart illustrating a training method for an actor critic network provided in one embodiment of this application; Figure 4 for Figure 3 A detailed flowchart of step 23; Figure 5 A schematic diagram of the functional modules of a cyberbullying detection system based on deep learning and value decision-making, provided in an embodiment of this application; Figure 6This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] This application provides a cyberbullying detection method based on deep learning and value decision-making, which can be applied to, for example... Figure 1 In the application environment shown, terminal 101 communicates with server 102 via a network. A data storage system can store the data that server 102 needs to process. The data storage system can be set up independently, integrated into server 102, or placed in the cloud or on another server. Terminal 101 can send raw text data to server 102. Server 102 obtains the raw text data to be detected, performs preprocessing to obtain a text identifier sequence, and then inputs the text identifier sequence into a fine-tuned pre-trained language model to extract the global semantic feature vector of the raw text data. This global semantic feature vector is then input into a multi-channel convolutional neural network to obtain a fused feature vector, which is then input into a trained actor / critic network to output a predicted probability of cyberbullying. Finally, based on the predicted probability, it is determined whether the raw text data constitutes cyberbullying. Server 102 can feed back the obtained predicted probability to terminal 101. In addition, in some embodiments, the prediction probability can also be implemented by the server 102 or the terminal 101 alone. For example, the terminal 101 can directly target the raw text data to be detected, or the server 102 can obtain the raw text data to be detected from the data storage system.

[0022] The terminal 101 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 102 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0023] In one exemplary embodiment, such as Figure 2 As shown, a method for detecting cyberbullying based on deep learning and value decision-making is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 We will use server 102 as an example to illustrate this.

[0024] like Figure 2 As shown, a cyberbullying detection method based on deep learning and value decision-making includes the following steps: S11, Obtain the original text data to be detected; S12, preprocess the original text data to obtain a sequence of text identifiers; S13, input the text identifier sequence into the fine-tuned pre-trained language model, and use the fine-tuned pre-trained language model to extract the global semantic feature vector of the original text data; S14, the global semantic feature vector is input into the multi-channel convolutional neural network, and local features at different scales are extracted by multiple parallel convolutional layers in the multi-channel convolutional neural network, and the local features at different scales are fused to obtain the fused feature vector; S15, the fused feature vector is input into the trained actor critic network, and the actor network in the trained actor critic network outputs the original text data as the predicted probability of cyberbullying; S16, Based on the predicted probability, determine whether the original text data constitutes cyberbullying.

[0025] By implementing steps 11 to 16 above, the acquisition and preprocessing of the original text data in steps 11 and 12 solves the problem that noise in the original social media text data leads to performance degradation when directly inputting it into the model. This standardizes the input data and provides a high-quality, standardized input sequence for subsequent advanced feature extraction. Step 13, based on the global semantic feature extraction of a pre-trained language model, addresses the difficulty of existing cyberbullying detection methods in deeply understanding the semantic context of text and capturing long-distance dependencies. This achieves the capture of global deep semantic information such as insulting intent and aggressive context, improving semantic understanding capabilities. Step 14, based on multi-channel convolutional neural network-based multi-scale local feature fusion, solves the problem of a single pre-trained model's inability to capture key local information. To address the issue of insufficient feature extraction capabilities, key local information such as specific abusive words and phrase combinations is incorporated. Furthermore, the combination of steps 14 and 13 solves the problem of single feature extraction in existing methods, achieving complementary fusion of global contextual semantics and multi-scale local key patterns, thus enhancing the model's ability to distinguish complex bullying expressions. The combination of steps 15 and 16 is also addressed. Step S15 uses a trained actor network, which has internalized a dynamic optimization strategy based on value assessment. Therefore, it can output a prediction probability optimized by value decision. Finally, step S16 determines whether the original text data is cyberbullying based on the optimized prediction probability, ultimately improving the overall detection performance. This enables more accurate identification of cyberbullying behavior and significantly reduces the false positive rate, especially improving adaptability to ambiguous language scenarios and enhancing generalization ability.

[0026] In practice, the raw text data in step 11 refers to user-generated short texts collected from social media platforms or online communities, which are the direct targets of cyberbullying detection. Raw text data mainly includes post text, user comments, and private message conversations, typically obtained through platform compliance interfaces or public datasets. To adapt to subsequent data processing, the raw data needs to be stored in a structured format, specifically as a CSV structured file. Each data entry should contain at least two core fields: "Text Content" and "True Label." The "Text Content" field stores the raw string to be detected, while the "True Label" field is used during the training phase to identify whether the sample constitutes cyberbullying (e.g., bullying is labeled as 1, non-bullying as 0). During the application detection phase, this "True Label" field is empty or does not exist.

[0027] Understandably, before being input into the system, the raw text data is in the form of an unprocessed string sequence, which may contain noise information unrelated to semantics, such as "@" user mention tags, redundant spaces, special emojis, or line breaks. This noise can interfere with the model's understanding of the core semantics. Therefore, the first step in the preprocessing stage of step 12 is to clean the text of semantically irrelevant noise information. Next, after text cleaning, the text can be batch-encoded using a tokenizer, converting it into a sequence of tokens that the pre-trained language model can recognize. The text is then processed by the tokenizer to ultimately generate input_ids (word embedding identifiers) and attention_masks. Therefore, after performing the above series of preprocessing steps on the raw text data, the resulting text identifier sequence can be represented as input data that the pre-trained language model can recognize, such as input_ids and attention_masks.

[0028] In practice, the pre-trained language model in step 13 can be a RoBERTa pre-trained language model. The RoBERTa pre-trained language model is a RoBERTa-base model based on the Transformer architecture, and its overall structure consists of an embedding layer and multiple Transformer encoder layers.

[0029] Specifically, the embedding layer includes Token Embedding, Position Embedding, and Segment Embedding. The embedding layer is used to convert the input_ids of the text identifier sequence into a word embedding tensor and inject positional information. For example, the input_ids mentioned above are converted into 768-dimensional word vectors.

[0030] Meanwhile, the aforementioned attention mask is passed as an independent input tensor along with the word embedding tensor to the subsequent Transformer encoder. This mask plays a role in the self-attention computation of each encoder layer, used to mask the positions representing padding tokens, ensuring that the model's attention mechanism focuses only on the real text content, thereby guaranteeing the correctness of batch computation.

[0031] The Transformer encoder layer can have 12 layers. Each Transformer encoder layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer, with 12 parallel attention heads within each sublayer. The multi-head self-attention sublayer is used to capture the semantic relationships between different tokens in the text, and the feedforward neural network sublayer is used to perform non-linear transformations on the attention output. As the word embedding tensor input passes through the 12 Transformer encoder layers sequentially, each layer calculates the attention weights between tokens and updates the word vectors through the multi-head self-attention mechanism. The Transformer encoder layer finally outputs a last hidden state tensor, which serves as a global semantic representation, containing feature vectors with rich deep semantic information. Subsequently, the output layer of the RoBERTa-base model based on the Transformer architecture performs dimensionality permutation on the last hidden state tensor to obtain a feature tensor, which serves as the input to the multi-channel convolutional neural network in step 14.

[0032] In specific implementation, the multichannel convolutional neural network (MCNN) in step 14 includes three parallel one-dimensional convolutional layers, three adaptive max pooling layers, and one fully connected layer.

[0033] The three one-dimensional convolutional layers are conv1, conv2, and conv3. conv1 has a kernel size of 2 to capture fine-grained local features; conv2 has a kernel size of 4 to capture medium-grained local features; and conv3 has a kernel size of 6 to capture coarse-grained local features. Each one-dimensional convolutional layer is followed by an adaptive one-dimensional max-pooling layer, which pools the convolutional outputs of different lengths into vectors with a fixed value of 1, thereby extracting the most significant feature response in each feature channel and achieving feature dimensionality reduction. The feature vectors output from the three pooling branches are concatenated along the channel dimension to form a joint feature vector that integrates multi-scale information. This concatenated joint feature vector is then passed through a fully connected layer to map its dimension back to the same dimension as the input feature vector, such as 768 dimensions, to achieve dimensional alignment and deep fusion with the global semantic feature vector. The output of this fully connected layer is the fused feature vector, which retains the deep semantic information from the pre-trained model while enhancing multi-scale local discriminative features. This fused feature vector will serve as input to the actor critic network in step 15.

[0034] In practice, the actor-critic network consists of an actor network and a critic network. The predicted probability is output through the trained actor network. The actor network and the critic network share the same fused feature vector input, but their division of labor and data processing objectives are different. In the application detection stage, only the trained actor network is used, and the critic network does not participate.

[0035] Specifically, the actor network has a sequential structure, which includes: a fully connected compression layer, a ReLU activation function layer, a linear mapping layer, and a sigmoid output layer. The last layer uses the sigmoid activation function to map the result of the linear transformation into a scalar value in the interval (0,1). This value represents the predicted probability (π(a|s)) that the original text data is cyberbullying, characterizing the confidence level that the current input text is classified as cyberbullying. For example, the closer the probability value is to 1, the more confident the model is that the text represents bullying behavior; the closer it is to 0, the more confident it is not bullying.

[0036] Understandably, the predicted probability values ​​output by the aforementioned actor network are the direct and sole basis for the final binary classification decision in subsequent step S16. Since the aforementioned actor critic network is optimized and trained using a value decision mechanism, its output probabilities exhibit higher calibration and reliability compared to traditional static classifiers when dealing with complex texts such as those that are ambiguous or satirical.

[0037] In one exemplary embodiment, such as Figure 3 As shown, a training method for an actor critic network is provided, the specific steps of which include: S21, simultaneously input the fused feature vector of the training samples into the actor critic network with unoptimized parameters, output the predicted probability of the training samples through the actor network, and output the state evaluation value of the training samples in the current predicted state through the critic network. S22, using the true labels of the training samples as reward signals and utilizing state value assessment, calculates the advantage value; S23, based on the advantage value and the predicted probability of the training samples, optimize the parameters of the actor network and the critic network whose parameters have not been optimized.

[0038] Specifically, the training samples in step 21 also refer to short texts generated by users collected from social media platforms or online communities, and the format conversion and preprocessing of the training samples are the same as described above, so they will not be repeated here. Each piece of data in the training samples contains at least two core fields: "text content" and "real label". The "text content" field stores the original string to be detected, while the "real label" field is used to identify whether the sample is cyberbullying, such as bullying being labeled as 1 and non-bullying as 0. This "real label" field exists in the aromatherapy stage.

[0039] Specifically, the actor network in the unoptimized actor-critic network is the same as described above and will not be repeated here. The critic network also has a sequential structure, but its design and output objective differ from the actor network. The critic network is an evaluation module that runs parallel to the actor network and is dedicated to outputting state values. The structure of the critic network consists of: a fully connected compression layer, a ReLU activation function layer, and a linear mapping layer. The linear mapping layer maps the features to the final scalar output value, i.e., the state evaluation value.

[0040] Specifically, in step 22 above, the advantage value is calculated using the formula A(s,a)=R(s,a)-V(s), where A(s,a) represents the advantage value, R(s,a) represents the reward signal, and V(s) represents the state value assessment.

[0041] Specifically, such as Figure 4 As shown, step 23 above specifically includes: S231, Construct the loss function of the actor network based on the advantage value and the prediction probability of the training samples; S232, constructing the loss function of the critic network based on the advantage value; S233, construct the total loss function by weighted sum of the loss functions of the actor network and the critic network; S234 uses the total loss function to update the parameters of the actor network and the critic network.

[0042] The formula for constructing the loss function of the actor network is L. actor =-log(π(a|s))×A(s,a), where, L actor Let L be the loss function of the actor network, where π(a|s) represents the sample prediction probability and A(s,a) represents the advantage value; the loss function of the critic network is constructed using the formula L. critic =[A(s,a)] 2 In the formula, L critic Let A(s,a) be the loss function of the commentator network, and let A(s,a) be the advantage value.

[0043] By implementing steps 21 to 23 above, and combining steps 21 and 22, the real labels of the training samples are used as reward signals, which solves the problem of rigid decision-making logic in classification models under the traditional supervised learning training paradigm. Then, by introducing the critic network to output independent state value assessments and combining them with the reward signals to calculate the advantage value, a dynamic and value-based gradient signal is provided for the optimization of the actor network. This enables the actor network to not only pursue correct predictions during the learning process, but also to learn to make decisions with higher value. Through the sub-steps in step 23, the actor network loss function is constructed, and the advantage value is used as a weight to apply to the policy gradient. This drives the actor network to significantly increase the probability of decision actions that bring positive advantage values, while reducing the probability of decision actions that lead to negative advantage values, thus achieving targeted policy improvement. The critic network loss function is constructed, which drives the critic network to continuously update its parameters so that its state value assessment is as close as possible to the real reward, thereby improving the accuracy of its assessment. The parameters of the two networks are jointly updated using a weighted total loss function, realizing the synergistic optimization of the two. Accurate critic assessment can provide the actor network with more reliable advantage value signals, while the improvement of the actor network's policy, in turn, provides the critic network with richer state-value data pairs under the new policy for learning. The two form a virtuous cycle and converge to a better equilibrium point. Therefore, by introducing a critic network, dominance calculation, and a dual-network collaborative optimization mechanism based on this, the dynamic decision-making of reinforcement learning is integrated into the training of the text classification model. The trained actor network not only has high classification accuracy, but also has enhanced decision robustness and adaptability to cope with complex language scenarios such as ambiguity and irony. Ultimately, it achieves an improvement in overall detection performance, can more accurately identify cyberbullying behavior, and significantly reduces the false positive rate, especially improving adaptability to ambiguous language scenarios.

[0044] In one exemplary embodiment, such as Figure 5 As shown, a cyberbullying detection system based on deep learning and value decision-making is provided, including: The acquisition module 201 is configured to acquire the raw text data to be detected. Preprocessing module 202 is configured to preprocess the original text data to obtain a sequence of text identifiers; The feature vector extraction module 203 is configured to input the text identifier sequence into a fine-tuned pre-trained language model and use the fine-tuned pre-trained language model to extract the global semantic feature vector of the original text data. The feature fusion module 204 is configured to input the global semantic feature vector into a multi-channel convolutional neural network, extract local features of different scales through multiple parallel convolutional layers in the multi-channel convolutional neural network, and fuse the local features of different scales to obtain a fused feature vector. The prediction probability generation module 205 is configured to input the fused feature vector into a trained actor critic network, and output the original text data as the prediction probability of cyberbullying through the actor network in the trained actor critic network; The bullying determination module 206 is configured to determine whether the original text data constitutes cyberbullying based on the predicted probability.

[0045] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6 As shown. The computer device includes a processor, memory, input / output interfaces (I / O), and communication interfaces. The memory of the computer device includes non-volatile storage media and internal memory. The database of the computer device is used to store the raw text data to be detected. When the computer program is executed by the processor, it implements a cyberbullying detection method based on deep learning and value decision-making.

[0046] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0047] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0048] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0049] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0050] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Furthermore, any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory.

[0051] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0052] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A cyberbullying detection method based on deep learning and value decision, characterized in that, The method comprises the following steps: obtaining original text data to be detected; preprocessing the original text data to obtain a text identifier sequence; inputting the text identifier sequence into a fine-tuned pre-trained language model, and extracting a global semantic feature vector of the original text data by using the fine-tuned pre-trained language model; inputting the global semantic feature vector into a multi-channel convolutional neural network, extracting local features of different scales by a plurality of parallel convolutional layers in the multi-channel convolutional neural network respectively, and fusing the local features of different scales to obtain a fused feature vector; inputting the fused feature vector into a trained actor-critic network, and outputting a prediction probability of the original text data being cyberbullying by an actor network in the trained actor-critic network; determining whether the original text data is cyberbullying according to the prediction probability. 2.The deep learning and value decision based cyberbullying detection method of claim 1, wherein, The training process of the trained actor-critic network comprises the following steps: inputting the fused feature vector of a training sample into a parameter-unoptimized actor-critic network, outputting a prediction probability of the training sample by the actor network, and outputting a state evaluation value of the training sample in a current prediction state by a critic network; taking a true label of the training sample as a reward signal, and calculating an advantage value by using the state value evaluation; optimizing parameters of the parameter-unoptimized actor network and the critic network based on the advantage value and the prediction probability of the training sample. 3.The deep learning and value decision based cyberbullying detection method of claim 2, wherein, The optimization of the parameters of the parameter-unoptimized actor network and the critic network based on the advantage value and the prediction probability of the training sample comprises the following steps: constructing a loss function of the actor network based on the advantage value and the prediction probability of the training sample; constructing a loss function of the critic network based on the advantage value; constructing a total loss function by a weighted sum of the loss function of the actor network and the loss function of the critic network; updating the parameters of the actor network and the critic network by using the total loss function.

4. The deep learning and value decision based cyberbullying detection method of claim 3, wherein, The loss function of the actor network is constructed as L actor =-log(π(a|s))×A(s,a), wherein L actor is the loss function of the actor network, π(a|s) is the sample prediction probability, and A(s,a) is the advantage value.

5. The deep learning and value decision based cyberbullying detection method of claim 3, wherein, The loss function of the critic network is constructed as L critic = [A(s, a)] 2 , wherein L critic is the loss function of the critic network, and A(s, a) is the advantage value.

6. The deep learning and value decision based cyberbullying detection method of claim 2, wherein, The advantage value is calculated by formula A(s, a) = R(s, a) - V(s), wherein A(s, a) is the advantage value, R(s, a) is the reward signal, and V(s) is the state value evaluation.

7. The deep learning and value decision based cyberbullying detection method of claim 1, wherein, The multi-channel convolutional neural network comprises three parallel convolutional layers with different convolution kernel sizes, which are used to extract local features of different scales respectively. 8.The deep learning and value decision based cyberbullying detection method of claim 1, wherein, The pre-trained language model is a RoBERTa pre-trained language model. 9.A cyberbullying detection system based on deep learning and value decision, characterized in that, The method comprises the following steps: an acquisition module configured to obtain original text data to be detected; a preprocessing module configured to preprocess the original text data to obtain a text identifier sequence; a feature vector extraction module configured to input the text identifier sequence into a fine-tuned pre-trained language model, and extract a global semantic feature vector of the original text data by using the fine-tuned pre-trained language model; The feature fusion module is configured to input the global semantic feature vector into a multi-channel convolutional neural network, extract local features of different scales through a plurality of parallel convolutional layers in the multi-channel convolutional neural network respectively, and fuse the local features of different scales to obtain a fused feature vector. The prediction probability generation module is configured to input the fused feature vector into a trained actor-critic network, and output a prediction probability of the original text data being cyberbullying through an actor network in the trained actor-critic network. The bullying determination module is configured to determine whether the original text data is cyberbullying according to the prediction probability.

10. A computer device comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the cyberbullying detection method based on deep learning and value decision in any one of claims 1-8.

Images