Multi-dimensional information dynamic fusion method and system suitable for rumor detection in early text information insufficient situation

Abstract

This application relates to the field of information security technology and provides a multi-dimensional information dynamic fusion method and system for rumor detection in scenarios where early textual information is insufficient. By mapping high-dimensional textual features and multiple low-dimensional non-textual features of social media text to a unified hidden space, key semantic features and non-textual features of the same target dimension are obtained. Based on the key semantic features, a gating vector of non-textual features is generated to regulate the non-textual features, resulting in non-textual regulated features. The representation sequence composed of key semantic features and non-textual regulated features is fused to generate a fused representation sequence. Text position fusion representations are extracted from the fused representation sequence through position index slicing. The classification feature vector generated by concatenating the global average pooling vector of the fused representation sequence with the text position fusion representation is used to perform a binary classification task on the social media text to obtain the classification result of the social media text.

Description

Technical Field

[0001] This application relates to the field of information security technology, and in particular to a multi-dimensional information dynamic fusion method and system for rumor detection applicable to scenarios where early textual information is insufficient. Background Technology

[0002] Currently, deep learning models based on multimodal data fusion have become the mainstream technical solution in the field of media rumor detection. These technologies aim to go beyond the limitations of single text analysis and improve detection performance by integrating various information such as text, images, user behavior, and comment interactions.

[0003] For example, by employing complex feature fusion networks and knowledge distillation techniques, a balance between detection supervision and model efficiency is achieved. By combining hierarchical gated interactive fusion with lightweight model transfer techniques, the dynamic gap relationships between multimodal features such as semantics, sentiment, emoticons, and internet slang are explicitly modeled, thus overcoming the performance bottlenecks of traditional single-path feature extraction and shallow fusion methods. Simultaneously, by introducing a knowledge distillation framework, knowledge transfer and parameter compression between the teacher model and the student network are achieved while preserving deep feature association patterns, balancing the contradiction between detection accuracy and computational efficiency, ultimately improving the accuracy, generalization ability, and practical deployment applicability of rumor detection. Summary of the Invention

[0004] The purpose of this application is to provide a multi-dimensional information dynamic fusion method and system for rumor detection in situations where early textual information is insufficient, so as to solve or alleviate the problems existing in the above-mentioned prior art.

[0005] To achieve the above objectives, this application provides the following technical solution: This application provides a multi-dimensional information dynamic fusion method for rumor detection in situations where early text information is insufficient, including: mapping the high-dimensional text features and multiple low-dimensional non-text features of the acquired social media text to a unified hidden space to obtain key semantic features and non-text features of the same target dimension, and generating a gating vector about the non-text features of the target dimension based on the key semantic features of the target dimension. By modulating the non-textual features of the target dimension using gating vectors, non-textual modulated features of the target dimension are obtained. A representation sequence is then formed by combining the key semantic features of the target dimension with the non-textual modulated features in a fixed order. Perform feature fusion to generate a fused representation sequence. ; By modulating the non-textual features of the target dimension using gating vectors, non-textual modulated features of the target dimension are obtained. A representation sequence is then formed by combining the key semantic features of the target dimension with the non-textual modulated features in a fixed order. Perform feature fusion to generate a fused representation sequence. ; Preferably, the acquired social media texts are semantically encoded using a pre-trained Chinese language model to obtain high-dimensional textual features. and high-dimensional text features The key semantic features of the target dimension are obtained by sequentially performing linear projection, layer normalization, GELU activation, and Dropout. .

[0006] Preferably, the standardized input sequence obtained by tokenizing the acquired social media text is truncated or padded to a preset length; Attention masks generated by truncating or padding the standardized input sequence to a preset length, along with the standardized input sequence, are encoded using a pre-trained Chinese language model. Pooled output vectors are then extracted from the encoding results to represent high-dimensional textual features of the social media text. .

[0007] Preferably, the acquired social media text is processed according to the input specifications of the Chinese pre-trained language model, with a special classification marker [CLS] added at the beginning of the text segment and a separator marker [SEP] added at the end to generate a standardized input sequence.

[0008] Preferably, non-textual features include sentiment features, user features, and engagement features; correspondingly, The acquired social media text is segmented into words, and scores are calculated for each word in the segmentation results based on basic sentiment word assignment rules, degree adverb adjustment mechanisms, and negation word inversion logic to obtain low-dimensional sentiment features of the social media text. ; and low-dimensional emotional characteristics Mapping to a unified latent space generates sentiment features for the target dimension. ; or, Extracting low-dimensional user features from acquired social media text. Participation characteristics The data is standardized and mapped to a unified latent space to generate user features for the target dimension. Participation characteristics .

[0009] Preferably, the sentiment features include at least the total number of sentiment words, the number of positive sentiment words, the number of negative sentiment words, the maximum sentiment intensity, and the variance of the sentiment score; And / or, User characteristics include at least the following: authentication status, number of posts, number of followers, number of comments, number of likes, number of reposts, number of comments, number of valid sub-comments, dissemination time, dissemination speed, account age, and authentication type code.

[0010] Preferably, multiple independent gating networks with identical structures but different parameters are used to separately encode the key semantic features of the target dimension. Mapped to gated intermediate representations of multiple different low-dimensional non-textual features; The gated intermediate representations of multiple low-dimensional non-text features are converted into multiple different gated vectors with values ​​ranging from 0 to 1. The multiple different gated vectors are then multiplied element-wise with the corresponding low-dimensional non-text features to obtain the corresponding non-text control features of the target dimension.

[0011] This embodiment also provides a multi-dimensional information dynamic fusion system suitable for rumor detection in scenarios where early text information is insufficient. The system uses any of the aforementioned multi-dimensional information dynamic fusion methods suitable for rumor detection in scenarios where early text information is insufficient to classify the acquired social media text. The system includes: The initial feature unit is configured to map the high-dimensional text features and multiple low-dimensional non-text features of the acquired social media text to a unified hidden space to obtain key semantic features and non-text features of the same target dimension, and generate a gating vector of non-text features of the target dimension based on the key semantic features of the target dimension. The feature modulation and fusion unit is configured to modulate the non-textual features of the target dimension through a gating vector to obtain the non-textual modulated features of the target dimension, and to combine the key semantic features of the target dimension with the non-textual modulated features in a fixed order to form a representation sequence. Perform feature fusion to generate a fused representation sequence for the target dimension. ; Text classification units, configured as fused representation sequences from the target dimension via position index slicing. Extracting text location fusion representation And according to the fusion representation sequence Global average pooling vector Fusion representation with text location The classification feature vector generated by splicing The task of binary classification is performed on social media texts to obtain the classification results.

[0012] Beneficial effects: The multi-dimensional information dynamic fusion method and system for rumor detection in scenarios with insufficient early text information provided in this application's embodiments map the high-dimensional text features and multiple low-dimensional non-text features of the acquired social media text to a unified hidden space to obtain key semantic features and non-text features of the same target dimension. A gating vector for the non-text features of the target dimension is generated based on the key semantic features of the target dimension. Furthermore, the non-text features of the target dimension are modulated using the gating vector to obtain modulated non-text features of the target dimension. Finally, a representation sequence is formed by combining the key semantic features and the modulated non-text features of the target dimension in a fixed order. Perform feature fusion to generate a fused representation sequence. Next, the fused representation sequence is sliced ​​using position indexes. Extracting text location fusion representation And according to the fusion representation sequence Global average pooling vector Fusion representation with text location The classification feature vector generated by splicing The task of binary classification is performed on social media texts to obtain the classification results.

[0013] In this way, by comprehensively utilizing multi-dimensional information such as textual content and non-textual content (emotional expression, user attributes, and dissemination participation) within the same framework, and through text-guided gating mechanisms and fusion coding capabilities, textual content, emotional expression, user attributes, and dissemination can be uniformly represented and fused into a model. Under conditions of insufficient information or in the early stages, this enables the classification and processing of rumors in the early stages of social media texts where textual information is insufficient. Attached Figure Description

[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. Wherein: Figure 1 This is a flowchart illustrating a multi-dimensional information dynamic fusion method for rumor detection in scenarios where early textual information is insufficient, according to some embodiments of this application. Figure 2 This is a schematic diagram of the structure of a multi-dimensional information dynamic fusion system for rumor detection in scenarios where early textual information is insufficient, according to some embodiments of this application. Detailed Implementation

[0015] The present application will now be described in detail with reference to the accompanying drawings and embodiments. Various examples are provided by way of explanation and not by way of limitation. In fact, those skilled in the art will understand that modifications and variations can be made to the present application without departing from the scope or spirit of the present application. For example, a feature shown or described as part of one embodiment may be used in another embodiment to produce yet another embodiment. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention should fall within the scope of protection of the embodiments of the present invention.

[0016] When using deep learning models based on multimodal data fusion to detect media rumors, existing models mainly have the following problems: (1) Insufficient use of features. Some methods mainly rely on text information for discrimination, and do not make sufficient use of information such as user attributes, emotional expression and participation in dissemination; (2) Insufficient interaction between multidimensional features. Some methods mainly adopt simple splicing or fixed weight fusion, which makes it difficult to depict the dynamic relationship between text information and non-text information; (3) Features from different sources have differences in scale, expression form and noise level. If they are directly fused, it is easy to affect the stability of model training and the final discrimination effect; (4) In the early rumor detection scenario where information is insufficient, the information available is relatively limited. Under the condition of limited information, it is still a problem to be solved to achieve joint modeling of text content, user attributes, emotional expression and participation in dissemination.

[0017] Based on this, this embodiment proposes a multi-dimensional information dynamic fusion method for rumor detection in situations where early text information is insufficient. On the basis of uniformly representing features from different sources, it achieves joint modeling of text features, sentiment features, user features, and engagement features of social media texts through text-guided gating and fusion coding. This allows for the construction, regulation, fusion, and classification of multi-dimensional features within the same framework.

[0018] like Figure 1 As shown, this multi-dimensional information dynamic fusion method for rumor detection, applicable to scenarios with insufficient early textual information, includes: Step S101: Map the high-dimensional text features and multiple low-dimensional non-text features of the acquired social media text to a unified hidden space to obtain key semantic features and non-text features of the same target dimension, and generate a gating vector for the non-text features of the target dimension based on the key semantic features of the target dimension.

[0019] In this embodiment, the text features of the acquired social media text will be... and multiple different low-dimensional non-textual features (sentiment features) User characteristics and participation characteristics Each of these is mapped to a unified hidden space, yielding key semantic features in the same target dimension. Non-textual features (sentiment features) User characteristics and participation characteristics ).

[0020] Among them, text features are used to characterize the semantic information of social media texts, sentiment features are used to characterize the emotional fluctuations in social media texts, user features are used to characterize the identity attributes and historical behavior information of the publisher (social users) of social media texts, and engagement features are used to characterize the degree of disagreement in comments, the speed of dissemination, and the level of audience participation in the dissemination process of social media texts.

[0021] For each social media text, construct high-dimensional text features. and multiple different low-dimensional non-textual features (sentiment features) User characteristics and participation characteristics Since text features and non-text features differ in dimensionality and numerical range, this embodiment uses high-dimensional text features. and low-dimensional emotional characteristics User characteristics and participation characteristics Emotional characteristics User characteristics and participation characteristics Mapped to a unified anomaly space, forming a unified representation that can be used for joint modeling.

[0022] Specifically, a Chinese pre-trained language model is used to semantically encode the acquired social media texts, thereby obtaining high-dimensional textual features of the social media texts. First, the acquired social media text is tokenized to obtain a standardized input sequence. Specifically, according to the input specifications of the Chinese pre-trained language model, a special classification marker [CLS] is added to the beginning of the text segment and a separator marker [SEP] is added to the end of the segment to form a standardized input sequence. Then, the standardized input sequence is truncated or padded to a preset length to generate an attention mask. Next, the standardized input sequence and the generated attention mask are encoded using the Chinese pre-trained language model to obtain the encoded output. Finally, the pooled output vector is extracted from the encoded output as the high-dimensional text features of the social media text. .

[0023] Chinese pre-trained language models can use BERT, RoBERTa, or their Chinese variants to perform contextual semantic modeling on the input text sequence, and follow the model's specifications:

[0024] High-dimensional text features generated The key semantic features of the target dimension are obtained by sequentially performing linear projection, layer normalization, GELU activation, and Dropout. .

[0025] Non-textual features include sentiment features, user features, and engagement features. In constructing sentiment features, the acquired social media text is segmented, and scores are calculated for each word in the segmentation results based on basic sentiment word assignment rules, degree adverb adjustment mechanisms, and negation word inversion logic, thus obtaining low-dimensional sentiment features of the social media text. In this embodiment, the sentiment feature is a five-dimensional vector, including at least the total number of sentiment words, the number of positive sentiment words, the number of negative sentiment words, the maximum sentiment intensity, and the variance of the sentiment score. This is used to obtain the low-dimensional sentiment features. Then, according to the model:

[0026] Low-dimensional emotional characteristics Mapping to a unified latent space generates sentiment features for the target dimension. .

[0027] In terms of user feature construction, low-dimensional user features corresponding to the users are extracted from the acquired social media text. The user characteristics include at least authentication status, number of posts, number of followers, number of comments, number of likes, number of reposts, number of comments, number of valid sub-comments, dissemination time, dissemination speed, account age, and authentication type code. Then, the low-dimensional user characteristics are... Standardize the process and follow the model:

[0028] User characteristics in low latitudes Mapping to a unified latent space to generate user features for the target dimension .

[0029] In terms of engagement feature construction, low-dimensional engagement features of corresponding users are extracted from the acquired social media text. The engagement features include at least the degree of opinion divergence, the speed of dissemination, and the number of comments. Here, the degree of opinion divergence is determined based on the range of the sentiment score of the comment text, the speed of dissemination characterizes how quickly information spreads, and the number of comments characterizes the degree of audience participation. Then, the low-dimensional engagement features are... Standardize the process and follow the model:

[0030] Participation characteristics of low latitudes Mapping to a unified latent space generates engagement features for the target dimension. .

[0031] In this embodiment, key semantic features of the same target dimension are obtained. Emotional characteristics User characteristics Participation characteristics Subsequently, non-textual features are processed using a gating network, enabling them to adaptively adjust under the guidance of textual semantics. Specifically, multiple independent gating networks with identical structures but non-shared parameters are used to process the key semantic features of the target dimension. The mapping is performed on gated intermediate representations of non-textual features at different low dimensions, and these gated intermediate representations are then converted into different gated vectors with values ​​ranging from 0 to 1.

[0032] Specifically, the gating networks include an emotion gating network, a user gating network, and an engagement gating network. These three networks have the same structure but do not share parameters. Each gating network includes a learnable linear mapping layer and a sigmoid activation function. The learnable linear mapping layer is used to encode key semantic features. The mapping is to the gated intermediate representation of the corresponding non-text branch. The Sigmoid activation function is used to convert the gated intermediate representation into a gated vector with values ​​between 0 and 1.

[0033] In a specific application scenario, define non-text branches ,but

[0034] In the formula, Indicates emotional branches, Indicates a user branch. This represents the participation branch. The gating vectors corresponding to different gating branches. for:

[0035] in, , indicating a non-text branch The corresponding gating vector; This represents the Sigmoid activation function. Indicates non-text branches The learnable weight matrix of the gated network, Indicates non-text branches Learnable bias vectors in a gated network. Here, different non-text branches Since the parameters of the gated networks are not shared, the learnable weight matrix and learnable bias vector of the gated networks for different non-text branches are obtained independently by the corresponding branches.

[0036] Step S102: Modulate the non-textual features of the target dimension using gating vectors to obtain the non-textual modified features of the target dimension, and combine the key semantic features of the target dimension with the non-textual modified features in a fixed order to form a representation sequence. Perform feature fusion to generate a fused representation sequence. .

[0037] In this embodiment, key semantic features are used. The sentiment features for the target dimension are generated by inputting them into gating networks for the sentiment branch, user branch, and engagement branch, respectively. User characteristics Participation characteristics gating vector Next, multiple different gating vectors are multiplied element-wise with the corresponding low-dimensional non-textual features to obtain the corresponding non-textual control features for the target dimension. Specifically, according to the model:

[0038] Sentimental features for each target dimension User characteristics Participation characteristics The process involves adjusting the data to generate corresponding non-textual features for the target dimension—the adjusted sentiment features. User characteristics and participation characteristics .

[0039] Therefore, based on the semantic information of social media text, targeted feature manipulation is applied to three non-textual branches: sentiment, users, and engagement. This preserves and enhances non-textual information more relevant to the current text analysis, while suppressing weakly relevant or noisy components. Furthermore, the corresponding non-textual features of the target dimension (sentiment features) are then manipulated. User characteristics and participation characteristics Key semantic features of the target dimension A sequence is formed in a fixed order. :

[0040] For each social media text, there is For batch sizes of The collection of social media texts includes: Furthermore, this will represent the sequence. The input is a Transformer fusion encoder to perform feature fusion and generate a fused representation sequence. :

[0041] For a single social media text, there are: For batch sizes of The collection of social media texts includes: Here, the Transformer fusion encoder consists of two Transformer coding layers, with four attention heads. The feedforward network hidden layer dimension is twice the dimension of the unified hidden space, and GELU activation and Dropout mechanisms are employed.

[0042] Compared to conventional feature direct concatenation or fixed-weight fusion methods, the gating adjustment mechanism in this embodiment does not set fixed fusion weights for each non-text feature, nor does it generate gating values ​​separately from the non-text features themselves. Instead, it uses key semantic features. As a unified input condition, dedicated gating vectors for sentiment, user, and engagement are generated respectively, enabling non-text features to undergo adaptive adjustment based on text semantic conditions before entering the Transformer fusion encoder.

[0043] Step S103: The fused representation sequence from the target dimension is obtained by slicing the position index. Extracting text location fusion representation And according to the fusion representation sequence Global average pooling vector Fusion representation with text location The classification feature vector generated by splicing The task of binary classification is performed on social media texts to obtain the classification results.

[0044] In this embodiment, a fused representation sequence is generated by manipulating non-textual features of text semantics and by jointly fusing textual and non-textual features through encoding. Further from the fusion representation sequence Extract the fusion representation corresponding to the text position. Specifically, the input representation sequence of the Transformer fusion encoder is arranged according to "key semantic features". Emotional characteristics User characteristics and participation characteristics The representation sequence is constructed in a fixed order, with text features at the beginning. Here, the fused representation sequence output by the Transformer fusion encoder is defined. (in The output vector at the first position is taken as the fused representation of the text position, i.e. In the case of batch input, Then through Obtain the fused representation corresponding to the text position ,and .

[0045] At the same time, for the fusion representation sequence Global average pooling is performed along the sequence number, resulting in the corresponding global average pooling vector. Specifically, according to the model:

[0046] Calculate the global average pooling vector Among them, the average pooling vector It is used to aggregate the overall information from text branches, sentiment branches, user branches, and engagement branches after merging. The fusion sequence represents the sequence. The Middle The fusion vector corresponds to each position. In this embodiment, the fusion representation sequence It contains four positions. By averaging the fusion vectors corresponding to the four positions, the global average pooling vector is obtained. .

[0047] Furthermore, the obtained text location representation is fused. With global average pooling vector To assemble, specifically, according to the model:

[0048] Text position fusion representation With global average pooling vector By concatenating the vectors, we obtain the classification feature vector. In this embodiment, text position fusion representation With global average pooling vector All are of the same dimension, and the concatenated classification feature vector The dimension is 512.

[0049] Subsequently, the classification feature vector Input a multilayer perceptron classifier to perform binary classification on social media text. The multilayer perceptron classifier includes a first fully connected mapping layer, a first Dropout layer, a first normalization layer, a first GELU activation layer, a second fully connected mapping layer, a second Dropout layer, a second normalization layer, a second GELU activation layer, and an output layer. The specific classification process is as follows:

[0050] The 512-dimensional classification feature vector is processed through the first fully connected mapping layer. The first layer maps the 512-dimensional hidden features to 256-dimensional hidden features, and the second fully connected mapping layer maps the 512-dimensional hidden features to 256-dimensional hidden features. The output layer maps the 256-dimensional hidden features to a 2-dimensional output. Dropout, layer normalization, and GELU activation function are set after the first two mapping layers, with the Dropout ratio being 0.5.

[0051] Subsequently, the binary logits obtained from the output layer are processed using Softmax to obtain the classification probability distribution of the social media text, namely:

[0052] in, This represents the probability distribution of social media text belonging to different categories.

[0053] In this embodiment, by using the fused representation sequence Extract the text position corresponding to And calculate the global average pooling vector. While preserving both the dominant semantic information of the text and multidimensional fusion of global information, a classification feature vector is constructed. The final classification result is output through a multilayer perceptron classifier. This fully utilizes the content of social media text, as well as multi-dimensional information such as user identity, audience feedback, and dissemination popularity. Through a text-guided gating mechanism combined with Transformer fusion encoding processing, a unified representation and fusion model of features from different sources is achieved, enabling early rumor detection in social media text.

[0054] like Figure 2 As shown, this embodiment also provides a multi-dimensional information dynamic fusion system for rumor detection in situations where early text information is insufficient. The system uses the multi-dimensional information dynamic fusion method for rumor detection in situations where early text information is insufficient, as described in any of the above embodiments, to classify the acquired social media text. The system includes: The initial feature unit 201 is configured to map the high-dimensional text features and multiple low-dimensional non-text features of the acquired social media text to a unified hidden space to obtain key semantic features and non-text features of the same target dimension, and generate a gating vector of non-text features of the target dimension based on the key semantic features of the target dimension. Feature modulation and fusion unit 202 is configured to modulate the non-textual features of the target dimension through a gating vector to obtain the non-textual modulated features of the target dimension, and to combine the key semantic features of the target dimension with the non-textual modulated features in a fixed order to form a representation sequence. Perform feature fusion to generate a fused representation sequence for the target dimension. ; Text classification unit 203 is configured to represent the sequence of fused representations from the target dimension via position index slicing. Extracting text location fusion representation And according to the fusion representation sequence Global average pooling vector Fusion representation with text location The classification feature vector generated by splicing The task of binary classification is performed on social media texts to obtain the classification results.

[0055] The multi-dimensional information dynamic fusion system for rumor detection in scenarios with insufficient early text information provided in this embodiment can realize the steps and processes of the multi-dimensional information dynamic fusion method for rumor detection in scenarios with insufficient early text information in any of the above embodiments, and achieve the same technical effect, which will not be repeated here.

[0056] In the description of this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0057] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A multi-dimensional information dynamic fusion method for rumor detection applicable to scenarios with insufficient early textual information, characterized in that, include: The high-dimensional text features and multiple low-dimensional non-text features of the acquired social media text are mapped to a unified hidden space to obtain key semantic features and non-text features of the same target dimension. Based on the key semantic features of the target dimension, a gating vector for the non-text features of the target dimension is generated. By modulating the non-textual features of the target dimension using gating vectors, non-textual modulated features of the target dimension are obtained. A representation sequence is then formed by combining the key semantic features of the target dimension with the non-textual modulated features in a fixed order. Perform feature fusion to generate a fused representation sequence. ; From the fused representation sequence by position index slicing Extracting text location fusion representation And according to the fusion representation sequence Global average pooling vector Fusion representation with text location The classification feature vector generated by concatenation The task of binary classification is performed on social media texts to obtain the classification results.

2. The method according to claim 1, characterized in that, Semantic encoding of acquired social media texts is performed using a pre-trained Chinese language model to obtain high-dimensional textual features of the social media texts. And for high-dimensional text features The key semantic features of the target dimension are obtained by sequentially performing linear projection, layer normalization, GELU activation, and Dropout. .

3. The method according to claim 2, characterized in that, The standardized input sequence obtained by tokenizing the acquired social media text is truncated or padded to a preset length. Attention masks generated by truncating or padding the standardized input sequence to a preset length, along with the standardized input sequence, are encoded using a pre-trained Chinese language model. Pooled output vectors are then extracted from the encoding results to represent high-dimensional textual features of the social media text. .

4. The method according to claim 3, characterized in that, According to the input specifications of the Chinese pre-trained language model, special classification markers [CLS] are added at the beginning of the text fragments and separator markers [SEP] are added at the end to generate a standardized input sequence.

5. The method according to claim 1, characterized in that, Non-textual features include sentiment features, user features, and engagement features; Correspondingly, The acquired social media text is segmented into words, and scores are calculated for each word in the segmentation results based on basic sentiment word assignment rules, degree adverb adjustment mechanisms, and negation word inversion logic to obtain low-dimensional sentiment features of the social media text. ; and low-dimensional emotional characteristics Mapping to a unified latent space generates sentiment features for the target dimension. ; or, Extracting low-dimensional user features from acquired social media text. Participation characteristics The data is standardized and mapped to a unified latent space to generate user features for the target dimension. Participation characteristics .

6. The method according to claim 4, characterized in that, The sentiment characteristics include at least the total number of sentiment words, the number of positive sentiment words, the number of negative sentiment words, the maximum sentiment intensity, and the variance of the sentiment score; And / or, User characteristics include at least the following: authentication status, number of posts, number of followers, number of comments, number of likes, number of reposts, number of comments, number of valid sub-comments, dissemination time, dissemination speed, account age, and authentication type code.

7. The method according to claim 1, characterized in that, By using multiple independent gating networks with identical structures but different parameters, the key semantic features of the target dimension are respectively... Mapped to gated intermediate representations of multiple different low-dimensional non-textual features; The gated intermediate representations of multiple low-dimensional non-text features are converted into multiple different gated vectors with values ​​ranging from 0 to 1. The multiple different gated vectors are then multiplied element-wise with the corresponding low-dimensional non-text features to obtain the corresponding non-text control features of the target dimension.

8. A multi-dimensional information dynamic fusion system for rumor detection applicable to scenarios with insufficient early textual information, characterized in that, The system classifies acquired social media texts using any of the multi-dimensional information dynamic fusion methods applicable to rumor detection in scenarios with insufficient early textual information, as described in claims 1-8. The system comprises: The initial feature unit is configured to map the high-dimensional text features and multiple low-dimensional non-text features of the acquired social media text to a unified hidden space to obtain key semantic features and non-text features of the same target dimension, and generate a gating vector of non-text features of the target dimension based on the key semantic features of the target dimension. The feature modulation and fusion unit is configured to modulate the non-textual features of the target dimension through a gating vector to obtain the non-textual modulated features of the target dimension, and to combine the key semantic features of the target dimension with the non-textual modulated features in a fixed order to form a representation sequence. Perform feature fusion to generate a fused representation sequence for the target dimension. ; Text classification units, configured as fused representation sequences from the target dimension via position index slicing. Extracting text location fusion representation And according to the fusion representation sequence Global average pooling vector Fusion representation with text location The classification feature vector generated by concatenation The task of binary classification is performed on social media texts to obtain the classification results.

Images