Social robot detection method based on published content and user relationship fusion

By integrating detection methods of user posting content and relationships on social media, using RoBERTa and R-GCN models to extract text and relationship representations, the shortcomings of feature extraction and network construction in the prior art are solved, and the accuracy of social robot recognition is improved and The ability to identify robots' cluster camouflage.

CN118965109APending Publication Date: 2024-11-15UESTC (SHENZHEN) ADVANCED RES INST
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411015991.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The existing social robot detection methods have problems such as labeling errors, data quality dependence, difficulty in obtaining traffic data, and inconsistent graph model with reality in the process of feature extraction and network construction, and ignore the correlation of published content and dissemination link information. .

Method used

Using a fusion detection method based on published content and user relationships, the text representation and relationship network representation are extracted by preprocessing user data, and using RoBERTa encoding and R-GCN model, combined with the full connection layer and the relationship graph convolution neural network, user text is generated. Representation and relational network representation are identified.

Benefits of technology

Improved the accuracy of social robot recognition, and improved 1% and 3% compared with traditional methods on PAN-19 and Twibot-20 datasets, and significantly improved the recognition capability on the relationship network graph, improving the robot population The ability to identify camouflage.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118965109A_ABST
    Figure CN118965109A_ABST
Patent Text Reader

Abstract

The invention discloses a social robot detection method based on published content and user relationship fusion. The method comprises the following steps: uploading user data; preprocessing the user data to obtain preprocessed text data, attribute data and relational data; performing text representation extraction by utilizing the preprocessed text data and attribute data to obtain a first vector, and performing relational network representation extraction by utilizing the preprocessed text data, attribute data and relational data to obtain a second vector; performing result identification according to the first vector and the second vector; wherein the first vector is a user text representation vector, and the second vector is a user relation network representation vector. According to the method, the social robot is identified according to the content information published by the user on the social media and the social relation graph of the user, and the identification accuracy is improved through analysis and utilization of different types of data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of social robot identification, and specifically relates to a social robot detection method based on the fusion of published content and user relationships. Background Art

[0002] Current social robot detection methods can be divided into user feature-based methods and network structure-based methods. The former has evolved from using traditional feature engineering analysis methods to today's text encoding training to extract feature structures. The latter traditional method mainly analyzes network traffic and builds network graphs to analyze abnormal nodes in the network. With the research on deep neural network structures, graph neural networks are gradually applied to social media robot detection, that is, using user nodes and user relationship networks to build social graphs and obtain prediction results through training.

[0003] Currently, there are social robot detection methods based on feature engineering. By analyzing and extracting features of different categories, feature sets are obtained for training. However, the extraction process of these feature sets is closely related to the quality of the data set. In most cases, due to the lack of real data, manual labeling is required. The characteristics of malicious accounts are a mixture of automatic and human behaviors, which are prone to labeling errors. In addition, text content information is an important feature component of social robot detection. Most detections encode the published content and train it through deep networks, ignoring the correlation between the published content information. In addition to the difficulty in obtaining traffic data, traditional social robot detection based on network traffic is relatively simple when building graph models, which is poorly consistent with reality, and the performance depends on the assumptions of the algorithm. Many methods based on graph neural networks ignore important information such as account publishing content, dissemination links, and the community to which the person belongs. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a social robot detection method based on the fusion of published content and user relationships, which aims to identify social robots according to the content information posted by users on social media and their social relationship graphs, and improve the recognition accuracy by analyzing and utilizing different types of data.

[0005] The technical solution adopted by the present invention is:

[0006] In a first aspect, the present application discloses a social robot detection method based on the fusion of published content and user relationships, comprising the following steps:

[0007] Upload user data;

[0008] Preprocessing the user data to obtain preprocessed text data, attribute data, and relationship data;

[0009] Using the preprocessed text data and attribute data to extract text representation to obtain a first vector, and using the preprocessed text data, attribute data and relationship data to extract relationship network representation to obtain a second vector;

[0010] Perform result identification according to the first vector and the second vector;

[0011] The first vector is a user text representation vector, and the second vector is a user relationship network representation vector.

[0012] As an optional technical solution, the preprocessing includes:

[0013] Acquire text data, attribute data and relationship data from the user data;

[0014] Segmenting the text data uniformly and encoding it using a first encoding method;

[0015] Classifying the attribute data into true or false data items, digital attribute items or position data items, encoding the true or false data items using the second encoding method, encoding the digital attribute items using the third encoding method, and encoding the position data items using the fourth encoding method;

[0016] Among them, the relationship data includes explicit relationship data and implicit relationship data, the explicit relationship data includes the user's follow-up relationship and friend relationship data, and the implicit relationship data contains tag co-occurrence, and the tag co-occurrence is extracted from the text data; the first encoding method is RoBerta encoding, the second encoding method is zero-one encoding, the third encoding method is z-score encoding, and the fourth encoding method is one-hot encoding.

[0017] As an optional technical solution, the text representation extraction includes:

[0018] Using the encoded text data, extracting the third vector and the fourth vector, and using the encoded attribute data, obtaining the fifth vector;

[0019] Establishing a fully connected layer according to the third vector, the fourth vector and the fifth vector to obtain the first vector;

[0020] Among them, the third vector is a user text information vector, the fourth vector is a semantic consistency representation vector, and the fifth vector is a user attribute representation vector.

[0021] As an optional technical solution, the extraction of the third vector includes the following steps:

[0022] Using the encoded text data, generate the forward hidden state sequence of the jth text and the backward hidden state sequence

[0023] According to the sequence and Aggregate the forward and backward results to form the jth text representation sequence and Then obtain the representation vector of the jth text

[0024] Generate a representation vector for each user's text After getting the collection

[0025] Using the collection A forward and a reverse sequence are generated, and the results of the forward sequence and the reverse sequence are connected to form the third vector.

[0026] As an optional technical solution, the extraction of the fourth vector includes the following steps:

[0027] Use the language training model to get the original attention weight matrix M i ;

[0028] Use the maximum pooling function to i Downsampling, reducing the matrix M i Get the matrix

[0029] For the matrix Expand to get the consistency representation vector d of each layer i :

[0030] For all the obtained consistency representation vectors d i The weighted function is used to combine and obtain the fourth vector.

[0031] As an optional technical solution, the relationship network representation extraction includes: based on the relationship graph convolutional neural network R-GCN, using the encoded attribute data and text data as user node features, performing node aggregation according to the node features and the relationship data, and outputting the second vector.

[0032] As an optional technical solution, the node aggregation includes:

[0033] Convert node features to get the initial hidden vector of the node

[0034] The vector After passing through the l-layer relational graph convolutional neural network R-GCN, the intermediate vector is obtained

[0035] The second vector is obtained using the intermediate vector.

[0036] In a second aspect, the present application also discloses a social robot detection device based on the fusion of published content and user relationships, including:

[0037] Data upload module, used to upload user data;

[0038] A preprocessing module, used to preprocess the user data to obtain preprocessed text data, attribute data and relationship data;

[0039] A vector extraction module, used to extract text representation using the preprocessed text data and attribute data to obtain a first vector, and also used to extract relational network representation using the preprocessed text data, attribute data and relational data to obtain a second vector;

[0040] An identification module, used for performing result identification according to the first vector and the second vector;

[0041] The first vector is a user text representation vector, and the second vector is a user relationship network representation vector.

[0042] As an optional technical solution, the preprocessing module includes:

[0043] A data acquisition unit, used to acquire text data, attribute data and relationship data from the user data;

[0044] A first encoding unit, used for uniformly segmenting the text data and encoding it using a first encoding method;

[0045] a second encoding unit, configured to classify the attribute data into true or false data items, digital attribute items or position data items, and encode the true or false data items using a second encoding method, encode the digital attribute items using a third encoding method, and encode the position data items using a fourth encoding method;

[0046] Among them, the relationship data includes explicit relationship data and implicit relationship data, the explicit relationship data includes the user's follow-up relationship and friend relationship data, and the implicit relationship data contains tag co-occurrence, and the tag co-occurrence is extracted from the text data; the first encoding method is RoBerta encoding, the second encoding method is zero-one encoding, the third encoding method is z-score encoding, and the fourth encoding method is one-hot encoding.

[0047] As an optional technical solution, the vector extraction module includes a text representation extraction module and a relationship network representation extraction module, and the text representation extraction module includes:

[0048] A vector extraction unit, used to extract the third vector and the fourth vector according to the encoded text data, and obtain the fifth vector using the encoded attribute data;

[0049] A fully connected unit, used to establish a fully connected layer according to the third vector, the fourth vector and the fifth vector to obtain the first vector;

[0050] Among them, the third vector is a user text information vector, the fourth vector is a semantic consistency representation vector, and the fifth vector is a user attribute representation vector.

[0051] As an optional technical solution, the relationship network representation extraction module includes:

[0052] A node feature generating unit, used to generate user node features according to the encoded attribute data and text data;

[0053] A node aggregation unit is used to perform node aggregation according to the node features and the relationship data, and output the second vector.

[0054] In a third aspect, the present application also discloses a computer-readable medium, comprising a computer program, which enables the computer to execute the method disclosed above when the computer program is executed on the computer.

[0055] The beneficial effects of the present invention are:

[0056] 1. Compared with the traditional social robot detection technology based on content features, by adding semantic consistency representation vectors, the models of Kudugunta and Yang et al. are compared on the experimental data sets PAN-19 and Twibot-20. Kudugunta et al. proposed a model that uses both tweet content and metadata to detect tweet-level robots based on a contextual long short-term memory network. During the detection process, a large labeled dataset suitable for a deep network is generated from a very small amount of labeled data, and context features are extracted from the metadata and used as auxiliary input to the long short-term memory deep network that processes tweet text. Yang et al. made improvements to the versatility and detection rate of detecting social robot types, and screened out 20 features from the metadata, reducing the detection time while ensuring good detection results. The method of this application was used to conduct experiments, and the experimental results were averaged over five experiments. The experimental results show that the accuracy of the content semantic consistency enhancement model of this application is 1% and 3% higher than the accuracy of the best performance model, respectively.

[0057] 2. In the relational network graph, a relational graph convolutional network model based on heterogeneous graph is proposed to solve the problem of robot cluster camouflage. This model has improvements in relational edge extraction and node feature selection compared with traditional graph-based detection methods. Specifically, by calculating information gain to retain node features with great influence, two types of relationships, explicit and implicit, are extracted to form edges in the heterogeneous graph. Using the graph neural network structure, robots with clustering phenomena and good at camouflage are identified. Five experiments were conducted using the method of this application, and the average accuracy of the experimental results was taken. Compared with traditional feature-based detection and other graph-based detection methods, the accuracy of the model of this application on the Twibot-20 and MGTAB datasets was improved by 1.6% and 1.3% respectively compared with the best performing model. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the process of social robot detection method based on the fusion of published content and user relationships in an embodiment.

[0059] Figure 2 Schematic diagram of the text representation extraction model in the embodiment.

[0060] Figure 3 Schematic diagram of the relational network representation extraction model in the embodiment.

[0061] Figure 4 Schematic diagram of the node aggregation process in the embodiment.

[0062] Figure 5 It is the data flow diagram of the overall detection system.

[0063] Figure 6 It is a schematic diagram of a social robot detection device based on the fusion of published content and user relationships. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0065] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0066] Example

[0067] like Figure 1 As shown, this embodiment discloses a social robot detection method based on the fusion of published content and user relationships, including the following steps:

[0068] Upload user data;

[0069] Preprocessing the user data to obtain preprocessed text data, attribute data, and relationship data;

[0070] Using the preprocessed text data and attribute data to extract text representation to obtain a first vector, and using the preprocessed text data, attribute data and relationship data to extract relationship network representation to obtain a second vector;

[0071] Perform result identification according to the first vector and the second vector;

[0072] The first vector is a user text representation vector, and the second vector is a user relationship network representation vector.

[0073] Specifically, the preprocessing results are stored, and the recognition results are stored; the user recognition results are fed back to the query system, and it is asked whether to conduct an accurate query, that is, to query whether a certain ID is a robot, and the results are fed back to the query interface through database interaction.

[0074] As an optional implementation, the preprocessing includes:

[0075] Acquire text data, attribute data and relationship data from the user data;

[0076] Segmenting the text data uniformly and encoding it using a first encoding method;

[0077] Classifying the attribute data into true or false data items, digital attribute items or position data items, encoding the true or false data items using the second encoding method, encoding the digital attribute items using the third encoding method, and encoding the position data items using the fourth encoding method;

[0078] Among them, the relationship data includes explicit relationship data and implicit relationship data, the explicit relationship data includes the user's follow-up relationship and friend relationship data, and the implicit relationship data contains tag co-occurrence, and the tag co-occurrence is extracted from the text data; the first encoding method is RoBerta encoding, the second encoding method is zero-one encoding, the third encoding method is z-score encoding, and the fourth encoding method is one-hot encoding.

[0079] As an optional implementation, the text representation extraction includes:

[0080] Using the encoded text data, extracting the third vector and the fourth vector, and using the encoded attribute data, obtaining the fifth vector;

[0081] Establishing a fully connected layer according to the third vector, the fourth vector and the fifth vector to obtain the first vector;

[0082] Among them, the third vector is a user text information vector, the fourth vector is a semantic consistency representation vector, and the fifth vector is a user attribute representation vector.

[0083] As an optional implementation manner, the extraction of the third vector includes the following steps:

[0084] Using the encoded text data, generate the forward hidden state sequence of the jth text and the backward hidden state sequence

[0085] According to the sequence and Aggregate the forward and backward results to form the jth text representation sequence and Then obtain the representation vector of the jth text

[0086] Generate a representation vector for each user's text After getting the collection

[0087] Using the collection A forward and a reverse sequence are generated, and the results of the forward sequence and the reverse sequence are connected to form the third vector.

[0088] As an optional implementation manner, the extracting of the fourth vector comprises the following steps:

[0089] Use the language training model to get the original attention weight matrix M i ;

[0090] Use the maximum pooling function to i Downsampling, reducing the matrix M i Get the matrix

[0091] For the matrix Expand to get the consistency representation vector d of each layer i :

[0092] For all the obtained consistency representation vectors d iThe weighted function is used to combine and obtain the fourth vector.

[0093] As an optional implementation, the relationship network representation extraction includes: based on the relationship graph convolutional neural network R-GCN, using the encoded attribute data and text data as user node features, performing node aggregation according to the node features and the relationship data, and outputting the second vector.

[0094] As an optional implementation manner, the node aggregation includes:

[0095] Convert node features to get the initial hidden vector of the node

[0096] The vector After passing through the l-layer relational graph convolutional neural network R-GCN, the intermediate vector is obtained

[0097] The second vector is obtained using the intermediate vector.

[0098] In order to better understand this embodiment, the method of this embodiment is roughly divided into the following stages for further explanation:

[0099] A. Data input stage: It is necessary to classify and process the multimodal data according to the data set transmitted into the input system, and to divide the user data into user attribute data (i.e. the attribute data mentioned above) and user text data (i.e. the text data mentioned above).

[0100] B. Data preprocessing stage: The divided data needs to be cleaned and noise removed; then the user attribute data is Z-score standardized, one-hot encoded or zero-one encoded, the user text data is uniformly segmented, and encoded using RoBERTa; in this stage, the user's relationship data can also be extracted. The relationship data includes explicit relationship data and implicit relationship data. The explicit relationship data includes the user's follow-up relationship, friend relationship and fan list data. The implicit relationship includes label co-occurrence, which is extracted from the text data.

[0101] C. Text representation extraction stage: Using the encoded user text data, based on the bidirectional LSTM and Attention mechanism, the user text information vector (i.e., the third vector) is obtained, and the attention weight of the text is extracted to obtain the semantic consistency representation vector (i.e., the fourth vector); the user attribute representation vector (i.e., the fifth vector) is obtained by training using a fully connected layer and a nonlinear activation function layer after classification and encoding of the user attribute data. Based on the obtained third, fourth, and fifth vectors, a fully connected layer is established to obtain the user text representation vector (i.e., the first vector).

[0102] D. Relationship graph construction stage: The encoded user attribute data and user text data are used as user node features. According to the explicit and implicit relationships among users, a relationship network is constructed, and the user relationship network representation vector (i.e., the second vector) is output based on the two-layer R-GCN.

[0103] E. Model detection stage: The first vector obtained in the text representation extraction stage and the second vector obtained in the relationship graph construction stage are connected and then the recognition result is obtained through Softmax.

[0104] F. Result display stage: Output the account detection results. At the same time, display the user attribute data, analyze the composition of their friends, and draw a relationship network diagram.

[0105] In order to better understand this embodiment, each of the above stages is further described below.

[0106] like Figure 2 As shown, it is a schematic diagram of the text representation extraction model. When extracting text representation, the user attribute representation vector (fifth vector) is trained using a fully connected layer and a nonlinear activation function layer after classification encoding of attribute data.

[0107] After the tweet text and user description are encoded by RoBERTa respectively, the text is trained with a mechanism combining LSTM and Self-Attention. At the same time, since the attention mechanism reflects the consistency and relevance of the text in time sequence, the attention weights are extracted after each round of training, and the matrix is ​​condensed by reducing the dimension through the maximum pooling technology, expanded and connected, and the semantic consistency representation vector (the fourth vector) is obtained through the linear layer.

[0108] The extraction of semantic consistency representation vectors requires the use of the LSTM algorithm as a language training model, and then a self-attention layer, so that the original attention matrix weight M can be obtained. i ∈R (T+1)×(T+1) , where i is the i-th user, R (T +1)×(T+1) is a T+1-dimensional user matrix. At this time, the dimension of the attention weight matrix is ​​large, and the maximum pooling function needs to be used to adjust the attention weight matrix M i Downsampling, shrinking the attention weight matrix to obtain

[0109]

[0110] Among them, K represents the hyperparameter of the matrix size. Then, the attention weight matrix Expand it to get the consistency representation vector d of each layer i :

[0111]

[0112] where θ sc is a learnable parameter. Finally, the attention vectors after the m-layer pooling are combined using a weighted function to obtain the semantic consistency representation vector r c , which is the fourth vector:

[0113]

[0114] Among them, W D and b D is a learnable parameter, σ(.) represents the activation function, and aggr(.) represents the weighting function.

[0115] For the jth text of the user, the RoBERTa encoding method is used to obtain the text embedding. In order to train the tweet text and description text, a bidirectional LSTM structure is used to generate the forward and backward hidden state sequences of the jth text respectively. and Use Q j Represents the length of the j-th text, and aggregates the forward and backward results to form the j-th text representation sequence:

[0116]

[0117] Where t represents a user tweet. The jth text after the LSTM and Attention network layers is represented as:

[0118]

[0119] in, Represents the weight of the i-th word in the j-th text. After generating a vector for each text, we get the set Generate a forward and reverse sequence, and then connect the results of the forward sequence and the reverse sequence to form a user text information vector That is the third vector.

[0120] The three extracted representation vectors are represented by semantic consistency vector r c , user text information vector And the user attribute representation vector are connected together, and the cross entropy is used as the loss function to obtain the user text representation vector, that is, the first vector.

[0121] like Figure 3 As shown, in the relational network representation extraction model, users are represented by nodes in the graph, the nodes store the user's features, and the relationships between users are represented by connecting lines between the nodes.

[0122] In order to select effective node features to represent users, information gain is used to measure the amount of information provided by a feature for correctly predicting the category. The information gain values ​​are sorted and features with higher information gain values ​​are selected to form a numerical feature set. In addition, the user's profile description and tweet text are encoded to obtain a text feature set. The two together constitute the user node feature.

[0123] In this method model, two types of relationships are used to construct the user relationship network. Among them, the explicit relationship includes the follow relationship and the friend relationship, and the implicit relationship includes the label co-occurrence. According to the three edge relationships, a relational graph convolutional network is constructed to obtain different edge representations for each node.

[0124] The aggregation process for a certain user node is as follows: Figure 4 When applying R-GCN to learn user representation, first, the user features are converted to obtain the initial hidden vector of the node in the graph, which is calculated as follows:

[0125]

[0126] Among them, W1 and b1 are learnable parameters, r i is the i-th user node representation, and D is the user embedding dimension. After l layers of R-GCN, the user representation is obtained:

[0127]

[0128] Where Θ is the projection matrix, N r (i) is the set of neighbor nodes of the relationship r of node i. In order to obtain a user representation vector that finally contains three types of edges, let the representation vector of each edge relationship pass through a neural network structure:

[0129]

[0130] Among them, W2 and b2 are learnable parameters, h i That is the relationship network representation vector of user i, that is, the second vector.

[0131] Use cross entropy as the loss function to calibrate the learning parameters:

[0132]

[0133] Among them, Y represents the set of annotated users, y i is the user’s real label, is the model's predicted value for the user, λ is a hyperparameter, and θ represents all learnable parameters in the model.

[0134] Finally, the results are predicted and stored. The first vector and the second vector obtained in the above process are passed through a fully connected hierarchy to obtain the detection result representation, and the result is stored in the corresponding database. The overall detection system data flow diagram is as follows: Figure 5 shown.

[0135] like Figure 6 As shown, in another embodiment, the present application also discloses a social robot detection device based on the fusion of published content and user relationships, including:

[0136] Data upload module, used to upload user data;

[0137] A preprocessing module, used to preprocess the user data to obtain preprocessed text data, attribute data and relationship data;

[0138] A vector extraction module, used to extract text representation using the preprocessed text data and attribute data to obtain a first vector, and also used to extract relational network representation using the preprocessed text data, attribute data and relational data to obtain a second vector;

[0139] An identification module, used for performing result identification according to the first vector and the second vector;

[0140] The first vector is a user text representation vector, and the second vector is a user relationship network representation vector.

[0141] As an optional implementation, the preprocessing module includes:

[0142] A data acquisition unit, used to acquire text data, attribute data and relationship data from the user data;

[0143] A first encoding unit, used for uniformly segmenting the text data and encoding it using a first encoding method;

[0144] a second encoding unit, configured to classify the attribute data into true or false data items, digital attribute items or position data items, and encode the true or false data items using a second encoding method, encode the digital attribute items using a third encoding method, and encode the position data items using a fourth encoding method;

[0145] Among them, the relationship data includes explicit relationship data and implicit relationship data, the explicit relationship data includes the user's follow-up relationship and friend relationship data, the implicit relationship data contains tag co-occurrence, and the tag co-occurrence is extracted from the text data. The first encoding method is RoBerta encoding, the second encoding method is zero-one encoding, the third encoding method is z-score encoding, and the fourth encoding method is one-hot encoding.

[0146] As an optional implementation, the vector extraction module includes a text representation extraction module and a relationship network representation extraction module, and the text representation extraction module includes:

[0147] The vector extraction unit is used to extract the third vector and the fourth vector according to the encoded text data, and obtain the fifth vector using the encoded attribute data; specifically, the extraction of the third vector and the fourth vector is detailed in the method in the above embodiment and will not be repeated here.

[0148] A fully connected unit, used to establish a fully connected layer according to the third vector, the fourth vector and the fifth vector to obtain the first vector;

[0149] Among them, the third vector is a user text information vector, the fourth vector is a semantic consistency representation vector, and the fifth vector is a user attribute representation vector.

[0150] As an optional implementation, the relationship network representation extraction module includes:

[0151] A node feature generating unit, used to generate user node features according to the encoded attribute data and text data;

[0152] The node aggregation unit is used to perform node aggregation according to the node features and the relationship data, and output the second vector. The specific generation process of the second vector is detailed in the method in the above embodiment and will not be repeated here.

[0153] In another embodiment, the present application further discloses a computer-readable medium, including a computer program. When the computer program is executed on a computer, the computer is enabled to execute the method disclosed in the above embodiment.

[0154] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device and computer-readable medium can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0155] Those skilled in the art will appreciate that the above modules can be distributed in the device according to the description of the embodiment, or can be changed accordingly to be in one or more devices different from the present embodiment. The modules of the above embodiments can be combined into one module, or can be further divided into multiple submodules (units).

[0156] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described here can be implemented by software, or by combining software with necessary hardware. Therefore, the technical solution according to the embodiment of the present application can be embodied in the form of a software product, which can be stored in a readable medium or on a network, and includes several instructions to enable an electronic device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) to execute the method according to the embodiment of the present application.

[0157] The above-mentioned readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0158] The above description is only a specific implementation mode of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. All technical solutions that fall within the scope defined by the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A social robot detection method based on the fusion of published content and user relationships, characterized in that: The following steps are involved: Upload user data; Preprocessing the user data to obtain preprocessed text data, attribute data, and relationship data; Using the preprocessed text data and attribute data to extract text representation to obtain a first vector, and using the preprocessed text data, attribute data and relationship data to extract relationship network representation to obtain a second vector; Perform result identification according to the first vector and the second vector; The first vector is a user text representation vector, and the second vector is a user relationship network representation vector.

2. The social robot detection method based on the fusion of published content and user relationships according to claim 1 is characterized in that: The pre-processing comprises: Acquire text data, attribute data and relationship data from the user data; Segmenting the text data uniformly and encoding it using a first encoding method; Classifying the attribute data into true or false data items, digital attribute items or position data items, encoding the true or false data items using the second encoding method, encoding the digital attribute items using the third encoding method, and encoding the position data items using the fourth encoding method; The relationship data includes explicit relationship data and implicit relationship data. The explicit relationship data includes the user's attention relationship and friend relationship data. The implicit relationship data includes tag co-occurrence, and the tag co-occurrence is extracted from the text data.

3. The social robot detection method based on the fusion of published content and user relationships according to claim 2 is characterized in that: The text representation extraction includes: Using the encoded text data, extracting the third vector and the fourth vector, and using the encoded attribute data, obtaining the fifth vector; Establishing a fully connected layer according to the third vector, the fourth vector and the fifth vector to obtain the first vector; Among them, the third vector is a user text information vector, the fourth vector is a semantic consistency representation vector, and the fifth vector is a user attribute representation vector.

4. The social robot detection method based on the fusion of published content and user relationships according to claim 3 is characterized in that: The extraction of the third vector comprises the following steps: Using the encoded text data, generate the forward hidden state sequence of the jth text and the backward hidden state sequence According to the sequence and Aggregate the forward and backward results to form the jth text representation sequence and Then obtain the representation vector of the jth text Generate a representation vector for each user's text After getting the collection Using the collection A forward and a reverse sequence are generated, and the results of the forward sequence and the reverse sequence are connected to form the third vector.

5. The social robot detection method based on the fusion of published content and user relationships according to claim 3 or 4, characterized in that: The extraction of the fourth vector comprises the following steps: Use the language training model to get the original attention weight matrix M i ; Use the maximum pooling function to i Downsampling, reducing the matrix M i Get the matrix For the matrix Expand to get the consistency representation vector d of each layer i : For all the obtained consistency representation vectors d i The weighted function is used to combine and obtain the fourth vector.

6. The social robot detection method based on the fusion of published content and user relationships according to claim 2 is characterized in that: The relational network representation extraction includes: based on the relational graph convolutional neural network R-GCN, using the encoded attribute data and text data as user node features, performing node aggregation according to the node features and the relational data, and outputting the second vector.

7. The social robot detection method based on the fusion of published content and user relationships according to claim 6 is characterized in that: The node aggregation includes: Convert node features to get the initial hidden vector of the node The vector After passing through the l-layer relational graph convolutional neural network R-GCN, the intermediate vector is obtained The second vector is obtained using the intermediate vector.

8. A social robot detection device based on the fusion of published content and user relationships, characterized in that: include: Data upload module, used to upload user data; A preprocessing module, used to preprocess the user data to obtain preprocessed text data, attribute data and relationship data; A vector extraction module, used to extract text representation using the preprocessed text data and attribute data to obtain a first vector, and also used to extract relational network representation using the preprocessed text data, attribute data and relational data to obtain a second vector; An identification module, used for performing result identification according to the first vector and the second vector; The first vector is a user text representation vector, and the second vector is a user relationship network representation vector.

9. The social robot detection device based on the fusion of published content and user relationships according to claim 8, characterized in that: The preprocessing module comprises: A data acquisition unit, used to acquire text data, attribute data and relationship data from the user data; A first encoding unit, used for uniformly segmenting the text data and encoding it using a first encoding method; a second encoding unit, configured to classify the attribute data into true or false data items, digital attribute items or position data items, and encode the true or false data items using a second encoding method, encode the digital attribute items using a third encoding method, and encode the position data items using a fourth encoding method; The relationship data includes explicit relationship data and implicit relationship data, the explicit relationship data includes the user's attention relationship and friend relationship data, and the implicit relationship data includes tag co-occurrence, which is extracted from the text data; The vector extraction module includes a text representation extraction module and a relationship network representation extraction module, and the text representation extraction module includes: A vector extraction unit, used to extract the third vector and the fourth vector according to the encoded text data, and obtain the fifth vector using the encoded attribute data; A fully connected unit, used to establish a fully connected layer according to the third vector, the fourth vector and the fifth vector to obtain the first vector; Wherein, the third vector is a user text information vector, the fourth vector is a semantic consistency representation vector, and the fifth vector is a user attribute representation vector; The relationship network representation extraction module includes: A node feature generating unit, used to generate user node features according to the encoded attribute data and text data; A node aggregation unit is used to perform node aggregation according to the node features and the relationship data, and output the second vector.

10. A computer-readable medium, characterized in that The method comprises a computer program, which, when executed on a computer, causes the computer to execute the method according to any one of claims 1 to 7.

Citation Information

Cited By

  • Social robot discovery method and system based on flow multi-modal feature fusion

    CN119760627A

  • Social robot discovery method and system based on traffic multi-modal feature fusion

    CN119760627B