Large model driven social platform cognitive confrontation perception method and system based on dynamic cooperative detection

Abstract

The invention relates to the technical field of network space supervision, in particular to a large-model-driven social platform cognitive adversarial perception method and system based on dynamic cooperative detection, and the method comprises the steps: constructing single social relation feature networks according to user behaviors in social network data, and combining the single social relation feature networks, obtaining a social relation feature composite network; coupling the social relation feature composite networks in different time slices to obtain community distribution on each time slice; according to the community distribution and the social text content in each community, utilizing a large language model to detect the narrative theme of each community; on the basis of the narrative themes of the communities and by utilizing a site detection model, obtaining sites of the communities for the specified event; and constructing a cognitive confrontation perception picture according to the cognitive confrontation process of each community on the full-time dimension of a specified event by taking the community as a perception object. According to the method, enterprises and institutions can be helped to better understand the cognitive situation of the network space, and support is provided for making a network purification decision.

Description

Technical Field

[0001] The present invention relates to the field of cyberspace supervision technology, and in particular to a large model-driven social platform cognitive adversarial perception method and system based on dynamic collaborative detection. Background Art

[0002] With the rapid development of network technology, the complexity of human society has increased dramatically, and the way of information transmission has been greatly expanded. In the current network society, information on social platforms will affect people's views, attitudes and beliefs, prompting people to form their own cognition. In this era, the construction of the cognitive domain is particularly important. It has risen to the third key area after the physical domain and the information domain. The cognitive domain mainly involves the guidance of group ideology and the shaping of social beliefs. It is an important area for maintaining social security, cultural security and ideological security. The current cognitive domain security is also facing unprecedented challenges. On the one hand, the spread of false information, online rumors and harmful ideologies may quickly distort public cognition. On the other hand, some institutions use cyberspace to conduct cognitive confrontation and cultural infiltration, undermining cultural security. Therefore, it has become particularly urgent to strengthen the security governance and capacity building of the cognitive domain.

[0003] However, at present, the perception and protection of cognitive domains are facing serious challenges. The main reasons are reflected in two aspects: 1. The cognitive domain is characterized by breadth and abstractness. On the one hand, the cognitive domain is a complex and extensive field, covering basic cognitive elements, specific cognitive activities, and cognitive change trends that change with the times. On the other hand, the cognitive domain is an abstract concept, which is a summary, induction and refinement of a series of specific cognitive scenarios, thinking and behaviors. Therefore, it is difficult to find a comprehensive and specific indicator system to perceive and analyze the offensive and defensive situation of the cognitive domain. 2. Cognitive confrontation is hidden. Some institutions may subtly influence the cognition and behavior of the target group without arousing public vigilance. Existing technical means are difficult to identify and analyze. There is an urgent need for a technology that can use cognitive confrontation collaborative groups to monitor public opinion, so as to maintain the healthy ecology of social media and the security of user information by analyzing and maintaining the tendency of online public opinion. Summary of the invention

[0004] To this end, the present invention provides a large-model driven social platform cognitive confrontation perception method and system based on dynamic collaborative detection, which is helpful to discover collaborative communities on social platforms, identify and respond to cognitive confrontation methods, perceive cognitive offensives, and help enterprises and institutions better understand the cognitive situation of cyberspace, facilitate support for making network purification decisions, and provide strong technical support for maintaining the health of social media and user safety.

[0005] According to the design scheme provided by the present invention, on the one hand, a large model-driven social platform cognitive confrontation perception method based on dynamic collaborative detection is provided, comprising:

[0006] Constructing a single social relationship feature network based on user behaviors in social network data, and compounding each single social relationship feature network to transform the social network data into a social relationship feature compound network;

[0007] Couple the composite networks of social relationship features in different time slices and obtain the community distribution on each time slice;

[0008] Based on the distribution of communities and the content of social texts in each community, the narrative themes of each community group are detected using a large language model;

[0009] Based on the narrative themes of each community group and using a pre-trained stance detection model, the stance detection model includes a Transformer encoder for obtaining word embedding representations of both the designated event and the group narrative theme, and a text-gated attention mechanism for matching the designated event and the group narrative theme using word embeddings to determine the emotional stance of the group narrative theme on the designated event;

[0010] Taking the community as the perception object, a cognitive confrontation perception picture is constructed according to the cognitive confrontation process of each community in the full time dimension of a specified event, and the full time dimension is from the beginning to the end of the event.

[0011] As a large model driven social platform cognitive confrontation perception method based on dynamic collaborative detection of the present invention, further, a single social relationship feature network is constructed according to user behavior in social network data, and each single social relationship feature network is compounded, including:

[0012] Taking individual user accounts in social platforms as the research objects, we construct the forwarding relationship feature network, the attention relationship feature network and the comment relationship feature network based on the user behavior of each individual user account;

[0013] According to the distribution weight of each single social relationship feature network in the social network, the forwarding relationship feature network, the attention relationship feature network and the comment relationship feature network are merged into an undirected authorized network, and the undirected authorized network is used as a social relationship feature composite network.

[0014] As a large model-driven social platform cognitive confrontation perception method based on dynamic collaboration detection of the present invention, further, the social relationship feature composite network in different time slices is coupled, and the community distribution on each time slice is obtained, including:

[0015] The characteristic network structure of social relations at each time point is used as a time slice, and the network data at the corresponding time point is coupled by continuous time slices to construct a multi-dimensional time composite graph, which is used to describe the structural characteristics of the community at different time points and the changes in the relationships between community members;

[0016] For multi-dimensional time composite graphs, the Leiden algorithm is used to separate nodes into disjoint communities, and nodes are moved recursively. The nodes and communities are merged by maximizing the quality function of each community to obtain the aggregated community distribution.

[0017] As a large model driven social platform cognitive confrontation perception method based on dynamic collaboration detection of the present invention, further, the Leiden algorithm is used to recursively merge and aggregate communities, including:

[0018] Initialize a queue using all nodes;

[0019] Remove the first node from the front of the queue and determine whether moving the node to the new community can increase the community quality function. If the community quality function increases, move the node to the new community and add neighbor nodes that do not belong to the new community and are not in the queue to the tail of the queue.

[0020] Go back to the step where you removed the node from the front of the queue and repeat moving the removed node to the new community until the queue is empty.

[0021] As a large model driven social platform cognitive confrontation perception method based on dynamic collaborative detection of the present invention, further, according to the community distribution and the social text content in each community and using a large language model to detect the narrative theme of each community group, it includes:

[0022] In each time slice, the language type of social texts in the community is detected, and the community is divided into two levels based on language type;

[0023] A large language model is used to perform topic analysis on the social text in each secondary community, and the topic analysis results of each secondary community are merged to obtain the narrative body of each overall community in each time slice. The large language model is TopicGPT, which uses LLMs to mine potential topics in a text collection through prompts.

[0024] As a large model driven social platform cognitive confrontation perception method based on dynamic collaborative detection of the present invention, further, based on the narrative theme of each community group and using a pre-trained stance detection model to obtain the stance of the community on a specified event, it includes:

[0025] The word fragment model is used to decompose social text and specified events into a series of word fragments. Each word fragment is mapped to a word space vector using the Transformer encoder, and positional encoding is used to add position information to each word fragment to obtain the word embedding vector representation of the corresponding encoding sequence output;

[0026] Use a gating structure to expand the attention of each word fragment to a fine-grained semantic level, use the interaction between hidden states and word vector representations to obtain sentence representations related to specified events in social texts, and use a multi-layer perceptron to obtain the number of emotion label categories and related probability distributions for specified events.

[0027] Statistical methods are used to aggregate the narrative themes of community groups, and the community group's stance on a given event is determined based on the frequency or weight of each emotional tag in the community group.

[0028] As a large model driven social platform cognitive confrontation perception method based on dynamic collaborative detection of the present invention, further, a cognitive confrontation perception picture is constructed according to the cognitive confrontation process of each community in the full time dimension of a specified event, including:

[0029] A multi-level cognitive perception model is constructed based on the community cognitive confrontation process. The multi-level cognitive perception model perceives the dynamic evolution of public opinion positions on designated events within and between communities from the intra-group level, inter-group level and global level. The evolution of public opinion positions on designated events within the community is measured by changes in community structure and community influence, the evolution of public opinion positions on designated events between communities is measured by the number of inter-group edges and the flow of community personnel, and the evolution of public opinion positions on designated events in the global environment of the social network is measured by changes in the total number of communities, the number of nodes in each community and the dominant community.

[0030] On the other hand, the present invention also provides a large model-driven social platform cognitive confrontation perception system based on dynamic collaborative detection, comprising: a data conversion module, a network coupling module, a topic detection module, a sentiment classification module and a confrontation perception module, wherein:

[0031] A data conversion module, used to construct a single social relationship feature network based on user behaviors in social network data, and to compound each single social relationship feature network to convert the social network data into a social relationship feature compound network;

[0032] The network coupling module is used to couple the composite networks of social relationship features in different time slices and obtain the community distribution on each time slice;

[0033] Theme detection module is used to detect the narrative theme of each community group based on the community distribution and the social text content in each community and using a large language model;

[0034] A sentiment classification module, which is used to obtain the community's stance on a specified event based on the narrative theme of each community group and using a pre-trained stance detection model, wherein the stance detection model includes a Transformer encoder for obtaining word embedding representations of both the specified event and the group narrative theme, and a text-gated attention mechanism for matching information between the specified event and the group narrative theme using word embeddings to determine the sentiment stance of the group narrative theme on the specified event;

[0035] The confrontation perception module is used to take the community as the perception object and construct a cognitive confrontation perception picture according to the cognitive confrontation process of each community in the full time dimension of the specified event, where the full time dimension is from the beginning to the end of the event.

[0036] Beneficial effects of the present invention:

[0037] The present invention constructs the dynamic change characteristics of the community by comprehensively analyzing the changes in indicators in three dimensions: the single community level, the community relationship level, and the global level, perceives cognitive confrontation from three dimensions, and constructs a comprehensive and three-dimensional cognitive confrontation situation; dynamic community detection is performed based on time series fusion, which effectively avoids the link effectiveness problem of traditional community detection algorithms, and can fully perceive the dynamic changes and complete life cycle of the community, providing strong support for downstream tasks; and for low-quality social platform text content with rich content, complex language, and little information, the large model is used to perform semantic understanding and information integration capabilities to deeply understand the theme core of the text; and the gated attention mechanism is used to focus on the most critical information for stance detection in the text, so that even in text content with low information content, specific events can be focused on, and the theme analysis is integrated with the large model to assist each other, and the stance of social platform users can be efficiently and accurately detected, which can help enterprises and institutions better understand the cognitive situation of cyberspace, facilitate the formulation of network purification decisions, provide support, and provide strong technical support for maintaining the healthy development of social media and user safety. It has good application prospects in the fields of cyberspace purification and public opinion monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the cognitive confrontation perception process of a large model-driven social platform based on dynamic collaboration detection in an embodiment;

[0039] Figure 2 This is a schematic diagram of the network construction process based on the relationship feature layer compound in the embodiment;

[0040] Figure 3 This is a schematic diagram of the process of building a relationship feature layer in the embodiment;

[0041] Figure 4 The community evolution detection process of time series fusion in the embodiment is intended;

[0042] Figure 5It is a schematic diagram of the time composite diagram in the embodiment;

[0043] Figure 6 It is a schematic diagram of the structure of the coupled time composite graph in the embodiment;

[0044] Figure 7 This is a schematic diagram of the Leiden algorithm process in the embodiment;

[0045] Figure 8 The Leiden algorithm effect of actual data at three time points in the embodiment is shown;

[0046] Fig. 9 This is a schematic diagram of narrative semantic enhancement of a low-quality sample group driven by a large model in the embodiment;

[0047] Fig.10 This is a schematic diagram of the TopicGPT topic modeling framework in the embodiment;

[0048] Fig.11 This is a schematic diagram of the subject analysis fusion process in the embodiment;

[0049] Fig.12 This is a schematic diagram of the topic fusion group stance detection process focusing on key information in the embodiment;

[0050] Fig.13 The working principle of the Transformer encoder with word fragments in the embodiment is shown;

[0051] Fig.14 This is a schematic diagram of the generalization process from individuals to communities in the embodiment;

[0052] Fig.15 This is a schematic diagram of the three-level architecture of the cognitive perception model in the embodiment;

[0053] Fig.16 The system architecture in the embodiment is shown in FIG. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.

[0055] Embodiments of the present invention, see Figure 1 As shown, a large model-driven social platform cognitive adversarial perception method based on dynamic collaborative detection is provided, including:

[0056] S101. Construct a single social relationship feature network based on user behaviors in social network data, and compound each single social relationship feature network to convert the social network data into a social relationship feature compound network.

[0057] Conventional network construction methods currently have difficulty processing complex datasets of social networks. Social network datasets have the following characteristics: (1) Huge data scale. Social network data usually contains millions or even billions of nodes and connections, which exceeds the processing capacity of conventional network construction methods; (2) Dynamicity. Social networks are dynamically changing. New relationships and nodes will continue to appear, while old relationships may disappear or change. This requires network construction methods to be able to adapt to this dynamic nature; (3) Diversity. There are various types of nodes and connections in social networks. Different relationships may have different meanings and weights. Conventional methods often cannot fully capture this diversity; (4) Community structure. There is a community structure in social networks, that is, some nodes are more closely connected than other nodes to form communities or groups. Conventional methods may not be able to effectively identify and utilize this structure; (5) Rich information. Social networks contain not only basic connection information, but also rich additional information, such as user-generated content, tags, timestamps, etc. This information is crucial for understanding and analyzing social networks, but conventional methods often cannot effectively utilize this information. Therefore, in order to effectively process social network data, it is necessary to develop new network construction methods that should be able to handle large-scale data, adapt to dynamic changes, capture diversity, identify community structures, and utilize rich additional information.

[0058] The actor in social network analysis is the basic unit of activity in a social network, and is usually set as a network node in network construction. Social network research requires selecting the most relevant social context and determining which organizations in this context cover these social network actors. The actor in social network analysis can be any social unit or entity, which can be an individual, organization, company, school, college, village, city, etc.

[0059] In the embodiment of this case, individual user accounts in the social platform are taken as the research objects, and the forwarding relationship feature network, the attention relationship feature network and the comment relationship feature network are constructed according to the user behavior of each individual user account; according to the distribution weight of each single social relationship feature network in the social network, the forwarding relationship feature network, the attention relationship feature network and the comment relationship feature network are merged into an undirected authorized network, and the undirected authorized network is used as the social relationship feature composite network.

[0060] like Figure 2As shown, in the specific network construction, the abstract universality of the relationship data of the social platform is extracted, and the important relationship features between users are selected to construct a single relationship feature network respectively; in order to fully and completely reflect the relationship structure of the social network, each relationship feature network is compounded according to a certain rule to form a multi-level, full-factor social network. On the social platform, the individual user accounts in a single social platform are the most accurate research objects that can be located and data can be accurately obtained. These single user accounts can reflect the network dynamics in real time, are the basic units of the collaborative group, and are also the most basic objects in cognitive attack and defense. Therefore, in the embodiment of this case, the single user account of a hot event is used as the selection range of the network node, and one network node corresponds to one user.

[0061] In order to improve accuracy and generalization ability, the selection of relationship feature layers is particularly important. In combination with specific social network related research and the common means of user relationships on various social platforms, in this embodiment, three relationship feature layers are selected: comment, forward, and follow. Examples of three relationship feature layers constructed based on these three relationships are as follows: Figure 3 As shown, assuming that graph G has N user nodes, then:

[0062] According to the forwarding of posts between these N users, a forwarding relationship network is established (if user A forwards a post to user B, there is an edge from node A to node B).

[0063] A follow-up relationship network is established based on the followers and user relationships of these N users (if user A follows user B, there is an edge from node A to node B).

[0064] A comment relationship network is established based on the comments on posts between these N users (if user A comments on user B, there is an edge from node A to node B).

[0065] Let the constant C represent the standard unit of the relational composite layer weight. According to the principle of data distribution: Analyze social network data to understand the distribution of various relation types in the network. If a certain relation type occupies the majority, it may be necessary to give it a higher weight.

[0066] Assume that a data set has N users and a total of S related posts sent by these N users. First, store the information of these N users as N nodes of the graph G, and then build a forwarding relationship network based on the forwarding of posts between these N users (if user A forwards a post to user B, there is an edge from node A to node B); build a follow-up relationship network based on the attention and user relationships of these N users (if user A follows user B, there is an edge from node A to node B); build a comment relationship network based on the comments on posts between these N users (if user A comments on user B, there is an edge from node A to node B).

[0067] Based on the multiple networks established above, they are finally merged into an undirected weighted network, named the relational feature composite network, assuming that the number of edges is M (M>=0) and the number of nodes is N (N>0).

[0068] The edge weight of the network that matches the relationship feature is set to W, and the edge weight of the forwarding relationship network is set to W 转发 , the edge weight of the attention network is set to W 关注 , the edge weight of the comment relationship network is set to W 评论 The relationship characteristics conform to the edge weight of the network W = W 转发 +W 关注 +W 评论 .

[0069] Therefore, the weight distribution can be expressed as follows:

[0070] W 关注 =C*M 关注 / M

[0071] W 转发 =C*M 转发 / M

[0072] W 评论 =C*M 评论 / M

[0073] W=W 转发 +W 关注 +W 评论

[0074] By compounding the relationship feature layers according to weight distribution, not only can a multi-level, full-factor social network model be formed to fully display the user relationship structure in the network, but this composite model can also provide strong support for subsequent downstream tasks.

[0075] S102: Couple the social relationship feature composite networks in different time slices, and obtain the community distribution on each time slice.

[0076] Community refers to the social characteristics manifested in the network topology. The entire network is composed of several communities. The connections between nodes in the same community are relatively close, and the connections between communities are relatively sparse. Communities in social network platforms are the basis of cognitive confrontation. The cognitive confrontation of social platforms is the competition and merger between different communities. However, due to the extreme complexity of social networks, the huge network topology data hides the specific distribution of communities; at the same time, the distribution and structure of communities change dynamically with the development of cognitive confrontation, the stimulation of external information and the passage of time. Therefore, an efficient community detection algorithm is needed, which can not only discover hidden collaborative communities based on the global network structure, but also detect the dynamic changes of communities with time series changes, perceive the evolution process of communities, and be used for analysis and perception of downstream tasks.

[0077] In commonly used coupled graphs, the same set of vertices but different sets of edges are usually considered. In this case, each node is identified by a single community, and there cannot be different communities in different graphs. This situation has great limitations and does not conform to the community analysis situation. In the actual network structure, communities are prone to abnormal situations, such as the existence of some new nodes at this point in time, or the disappearance of some nodes. Therefore, a more extensive and applicable method is needed to couple time-series graphs.

[0078] In the embodiment of this case, the characteristic network structure of social relations at each time point can be used as a time slice, and continuous time slices can be used to couple the network data of the corresponding time points to construct a multi-dimensional time composite graph, which is used to describe the structural characteristics of the community at different time points and the changes in the relationship between community members; for the multi-dimensional time composite graph, the Leiden algorithm is used to separate the nodes into non-overlapping communities, and the nodes are recursively moved, and the nodes and communities are merged by maximizing the quality function of each community to obtain the aggregated community distribution.

[0079] By integrating time information and applying efficient community detection algorithms, the evolution of communities is dynamically displayed based on accurate community division. This not only enables the system to cope with complex social networks and explore community structures with high granularity, but also efficiently displays the dynamic structural changes of communities, providing strong support for downstream tasks such as cognitive adversarial perception. Figure 4As shown in the figure, by coupling network data at different time points, a multi-dimensional time composite graph is constructed, which not only captures the structural characteristics of the community at different time points, but also reflects the dynamic interactions and relationship changes between community members. It can explore the community structure with high granularity, reveal the micro-dynamics within the community, and the macro-trends of interactions between communities. The Leiden algorithm can not only optimize the internal connection of the community through the two major steps of local node movement and network cohesion, but also deal with the problem of broken links within the community, thereby providing a more reasonable and accurate community division result.

[0080] In this embodiment, the network structure at a time point is called a slice. The essence of a slice is just a different graph, and there is no need to set the same vertices in any way. A time composite graph (Multiplex) is constructed from all slices, and then the composite graph is decomposed again in the layer corresponding to the slice to perform community detection.

[0081] Slice coupling, for example, two consecutive time windows, or just two different slices of a relation type. In theory, any two slices can be coupled, so again a graph can be used to represent the coupling relationship itself. Therefore, the nodes of this coupling graph are slices, and the (possibly weighted) links in the coupling graph represent the (possibly weighted) couplings between slices. Figure 5 As shown, this is an example with three different time slices, where slice 1 is coupled to slice 2, which is in turn coupled to slice 3.

[0082] The coupling graph consists of three nodes and a simple linear structure: 1--2--3. Using this approach, all slices are transformed into a temporal composite graph by putting all nodes in all slices into a large network. Therefore, each node is represented by a tuple (node, slice) in a sense. From this large network, only the edges between nodes defined in the same slice are taken, thus forming a separate layer. Finally, a layer is used to represent the coupling. Therefore, the number of layers obtained is equal to the number of slices, and an additional layer is needed to represent the coupling between slices. As Figure 6 As shown, the structure of each layer of the coupled time composite graph is displayed. The first layer is a coupling graph, and the subsequent layers are the network structures of each layer.

[0083] Multi-scale analysis based on time composite graphs can help us understand the changes in community structure in different time windows or under different conditions. In the process of dynamic network analysis, the membership and structure of the community may change over time. The coupling graph can capture this dynamic change and analyze how the community evolves over time. At the same time, this also enhances the robustness of the community detection algorithm. Through the coupling graph, the robustness of the community detection algorithm to noise and outliers can be enhanced because the coupling relationship can help the algorithm identify more stable community boundaries. The coupling graph can be used to optimize the performance of the community detection algorithm and increase the convergence speed of the algorithm.

[0084] Among them, the Leiden algorithm is used to recursively merge and aggregate communities, which can be designed to include:

[0085] Initialize a queue using all nodes;

[0086] Remove the first node from the front of the queue and determine whether moving the node to the new community can increase the community quality function. If the community quality function increases, move the node to the new community and add neighbor nodes that do not belong to the new community and are not in the queue to the tail of the queue.

[0087] Go back to the step where you removed the node from the front of the queue and repeat moving the removed node to the new community until the queue is empty.

[0088] like Figure 7 As shown in the figure, the algorithm starts from a, treating each node as a separate community, and then moves a single node from one community to another to improve the quality function and find the partition b. Then b is improved to c, and the community to which the node belongs does not change. The network is condensed based on the improved partition c. The community division of the condensed network is used as the partition b before improvement. For example, the red community in b is improved to two types of nodes in c, but both belong to one community, and the algorithm condenses to form two nodes. Then the algorithm continues to move the condensed network nodes, and at this time, f will no longer improve the partition result. Then repeat the above steps until no further improvement can be made. In the improvement stage, nodes do not necessarily merge with the community greedily, and it is not necessary to maximize the growth of the quality function. Instead, a node can merge with any community with an increased quality function, and this merger is randomly selected. The greater the increase in the quality function, the greater the probability that the community will be selected. The random order of community selection explores the community space more extensively. Only communities with a quality function greater than 0 are selected, which improves the algorithm speed and still finds the optimal partition of a set of nodes. The algorithm has a fast local movement process. After visiting all nodes once, the Leiden algorithm only visits the nodes whose neighbor nodes have changed, that is, they are moved to other communities.

[0089] The quality function can be expressed as:

[0090]

[0091] Where e is the number of edges in community c, n represents the number of nodes in community c, and the γ resolution parameter acts like a threshold: the density of communities should be at least γ, while the density between communities should be lower than γ. The higher the resolution, the more communities there are, and the lower the resolution, the fewer communities there are.

[0092] like Figure 8 As shown in the figure, on the real social network data set, we can see that the dynamic partitioning effect of the Leiden algorithm is relatively good. We can see obvious changes in the partitions, and the overall development law is in line with the law of network behavior. The overall community is relatively stable, but there are obvious mergers, expansions and changes between communities.

[0093] S103. Detect the narrative themes of each community group based on the community distribution and the social text content in each community and using a large language model.

[0094] Group narrative analysis refers to analyzing the cognition and values ​​of the entire group based on the stories, events, and opinions expressed by group members. In the field of cognitive confrontation perception, changes in the narrative themes of a specific community can often help perceive whether the relevant community has been infiltrated and guided. On social platforms, group narrative analysis technology is often interrelated with topic analysis technology, and group narratives are often a complex superposition of individual narrative themes. On social platforms, individual narrative themes are often presented by various types of text content posted by users on social platforms. Therefore, group narrative analysis must be suitable for the text attributes of social platforms, and needs to have the ability to generalize from a single individual to a group. At the same time, it involves the complexity of the expression methods of social platforms and requires the ability to understand multiple languages.

[0095] The text attributes of social platforms are different from the types of texts studied in conventional topic analysis. They are short, diverse, informal, emotional, and interactive. At the same time, the text content is rich and diverse, mainly divided into user-generated content (UGC), information sharing content, social interaction content, and cultural exchange content. In this case, Fig. 9 As shown in the figure, based on the large model driven by low-quality sample group narrative semantic enhancement, on the basis of detailed division of the community, the social platform text data of users within the community is comprehensively utilized to realize the generalization of topic analysis from individual to group level, which can not only deeply explore and understand the personal narratives of community members, but also reveal the collective cognition and behavior patterns of the entire community; through this multi-dimensional, cross-language, real-time dynamic analysis, it provides a solid data support and analysis foundation for subsequent downstream tasks such as perceptual cognitive confrontation.

[0096] Specifically, within each time slice, the language type of the social text in the community can be detected, and the community can be divided into two levels based on the language type; a large language model can be used to perform topic analysis on the social text in each secondary community, and the topic analysis results of each secondary community can be merged to obtain the narrative body of each overall community within each time slice. The large language model is TopicGPT, which uses LLMs to mine potential topics in a text collection through prompts.

[0097] Langdetect can help developers detect the language of text. It can detect 55 languages ​​including but not limited to Chinese, English, French, German, Spanish, etc., basically realizing the detection of various common languages. For languages ​​that are not yet included, it can be used to generate new language configuration files from Wikipedia summary database files or plain text by using the langdetect.jar tool to achieve detection of new languages. Langdetect determines the language of text by calculating the language-specific character distribution, usually with high accuracy. It is easy to use and has few dependencies. After installing the library, language detection can be achieved through simple function calls.

[0098] Therefore, in the embodiment of this case, the language type detection uses the langdetect library to detect the speech or comment text of each user, and the result is given by a list, which contains all the possible text language types and gives the confidence for each language type. In the text analysis process, texts with practical significance are often generated in the same language, and the langdetect library is very sensitive, and the detection confidence for such texts is almost always higher than 99%. Therefore, the texts with language type detection confidence greater than 99% are retained, and the rest are discarded. The number of texts of some language types is too small, and the results of analyzing them are usually more accidental, and may be generalized, and their topics cannot be accurately known. Therefore, a quantity threshold is set, and all texts with less than the threshold are discarded. And a regularization method is used to clean and purify text data. By defining a series of rules, those insignificant elements are accurately identified and eliminated, which significantly improves the quality and availability of text data, and provides a clearer, more accurate, and denoised text input for the large model, which not only helps the large model to more effectively identify and analyze the topics in the text, but also improves the accuracy and reliability of the final analysis results.

[0099] Due to the extensiveness of user content on social platforms, some texts involve extremely sensitive topics and are not suitable for online analysis by large models. Calling the API interface of a large model for analysis involves the issue of running speed, and it is necessary to ensure the goodness of the current network environment. It is not stable for a model that may perform analysis at any time. Using online Q&A or API interface mode, it is impossible to fine-tune the performance of the model, nor to optimize the results of topic analysis. In this embodiment, Fig.10 As shown in the figure, TopicGPT topic modeling is used to analyze topics. TopicGPT uses large language models (LLMs) to reveal potential topics in a text collection through prompts, which mainly includes two main stages: topic generation and topic assignment. In the topic generation stage, a set of topics is generated based on document samples and seed topics, and the topic set is further improved by merging similar topic pairs and removing uncommon topics. The topic assignment stage involves assigning new documents to the generated topics and providing citation support.

[0100] The topic generation phase uses large language models (LLMs) to analyze the input document collection and generate a set of descriptive topics. In this phase, new topics are first generated by prompting LLM based on the provided document samples and a set of seed topics. These topics are designed to capture the key concepts in the text data. Subsequently, the generated topics are refined, including merging semantically similar or repeated topic pairs, and removing minor topics that appear less frequently in the document collection to ensure the accuracy and diversity of the topic set. At the same time, the topic generation phase also creates a multi-level topic structure. Through the top-level topics generated in the refinement phase, the LLM is further prompted to generate more specific sub-topics at a deeper level, thereby constructing a rich and hierarchical topic system. This process not only improves the relevance and interpretability of the topics, but also enhances the flexibility and adaptability of topic modeling through iterative prompting and correction.

[0101] The purpose of the topic assignment phase is to effectively associate the pre-generated list of topics with the documents in the dataset. In this phase, the system first receives a new document and a set of generated topics, and then identifies and assigns one or more topics that best fit its content to the document by prompting a large language model (LLM). The output of the model includes the topic label assigned to the document, a document-specific description of the topic, and quoted text extracted from the document, which serves as evidence to support the assignment decision. This process not only enhances the interpretability and verifiability of the topic assignment, but also ensures the accuracy of the assignment results through a self-correction mechanism, allowing the model to identify and correct any malformed or low-quality topic assignments, thereby improving the quality and reliability of overall topic modeling.

[0102] TopicGPT can generate topics that are highly consistent with human classification, provide natural language descriptions, and enhance the interpretability of topics; at the same time, it allows users to customize topics through simple text prompts without retraining the model, showing excellent adaptability and user interactivity; in addition, TopicGPT ensures the accuracy and reliability of topic assignment through reference support and self-correction mechanisms, making topic modeling results easier to verify; it also supports multi-level topic structures, which can deeply explore complex topic relationships in text data, thereby providing a richer and more detailed perspective in various text analysis tasks. Compared with traditional topic models (such as LDA), the topics generated by TopicGPT are not only more consistent with human classification, but also provide natural language labels and descriptions, which greatly improves the interpretability of topics. At the same time, TopicGPT's framework is highly adaptable, allowing users to customize topics through simple text prompts without retraining the model. It also supports self-correction of topics and the construction of multi-level topics. The main reason for using TopicGPT is its human-centered design, which makes the topic modeling process more intuitive, flexible and easy to understand, and is particularly suitable for application scenarios that require high interpretability and customization.

[0103] A continuous learning mechanism can be established to enable the model to learn from new data and continuously optimize its topic analysis capabilities. To further enhance the accuracy of the analysis results, see Fig.11 As shown, thematic analysis was conducted on small groups of the same language type in each group, and then the results of the thematic analysis of these small groups were merged to obtain the final theme of the group.

[0104] S104. Based on the group narrative themes of each community and using a pre-trained stance detection model, the stance detection model includes a Transformer encoder for obtaining word embedding representations of both the specified event and the group narrative theme, and a text-gated attention mechanism for matching information between the specified event and the group narrative theme using word embeddings to determine the emotional stance of the group narrative theme on the specified event.

[0105] Community stance detection is to study the stance of the entire community towards a particular event. It is an effective combination of individual stances within the community, and involves in-depth interpretation and mining of textual content published by community members. Unlike group narrative analysis, community stance analysis focuses more on the stance and attitude of the current community towards a specific event, such as support, opposition, and neutrality.

[0106] The basis of stance detection is to understand the stance description of social platform text content on specific events. Stance description refers to the content of the user's stance and opinions on specific events in the text content of social platforms. The stance description contained in the text content of social platforms is often brief, event-specific, hidden, emotionally integrated and diverse in expression. Specifically: event-specificity means that stance descriptions on social platforms usually revolve around specific events or topics; hiddenness means that users' stance expressions on social platforms do not always directly point to specific events, and they will convey their stances through metaphors, sarcasm or implicit language; emotional integration means that stance descriptions in social platform texts are often closely related to emotions, and stance descriptions may have strong emotional colors, such as anger, joy, sadness, etc.; diversity of expression means that the expression forms of social platform texts are very diverse, including but not limited to text, emoticons, pictures, videos, etc.

[0107] See also Fig.12 As shown, in this embodiment, the Transformer structure and the gated attention mechanism are combined with specific event information and the gated attention mechanism to form matching information between tweets and objects, so as to more accurately judge the stance of tweets on specific objects. At the same time, this method uses sentiment classification as an auxiliary task, using sentiment information to assist stance discrimination.

[0108] Specifically, the word fragment model can be used to decompose social text and designated events into a series of word fragments, and the Transformer encoder can be used to map each word fragment to a word space vector. The position encoding is used to add position information to each word fragment to obtain the word embedding vector representation of the corresponding encoding sequence output; the gating structure is used to expand the attention of each word fragment to a fine-grained semantic level, and the interaction between the hidden state and the word vector representation is used to obtain the sentence representation related to the designated event in the social text, and the multi-layer perceptron is used to obtain the number of emotion label categories and related probability distribution of the designated event; and the statistical method is used to aggregate the narrative themes of the community group, and the community group’s position on the designated event is determined based on the frequency or weight of each emotion label in the community group.

[0109] Due to the characteristics of position descriptions in social platform texts, conventional Transformer encoders often ignore important and implicit information. In the position descriptions of social platforms, if the description of specific events is not introduced, the model often finds it difficult to identify the object of the position description. For example, the user's text content is "Very good! Support relevant policies", but if the specific event is not embedded, his specific position cannot be judged. Therefore, it is necessary to embed keywords for specific events such as "Benefiting farmers" into the text content.

[0110] Wordpiece is a text representation method based on subword units, which allows the model to handle out-of-date (OOV) and rare words. In traditional word segmentation methods, text is usually broken down into words in a fixed vocabulary, which is powerless when encountering unknown words. The wordpiece model breaks down words into smaller, meaningful units (word pieces), allowing the model to understand and process new or rare words composed of these word pieces.

[0111] The Transformer is chosen to process all elements in the sequence in parallel, improving computational efficiency. More importantly, its self-attention mechanism allows the model to consider all input elements in the sequence when generating each output element. This mechanism enables the model to capture long-distance dependencies, which is crucial for understanding the semantic structure in language.

[0112] In this embodiment, the model of the integrated word fragment model and the Transformer encoder can fully utilize the advantages of both to handle complex social platform position descriptions. The specific process is as follows Fig.13 As shown, it may include the following:

[0113] 1. Text segmentation: First, the input text is segmented through the word fragment model to decompose the text into a series of word fragments. These word fragments are the basic units that constitute the vocabulary. They can be single characters, roots, affixes, or any other meaningful language units.

[0114] 2. Word fragment embedding: Each word fragment is mapped to an embedding vector in a high-dimensional space. These vectors are obtained through pre-training and can capture the semantic information of the word fragment.

[0115] 3. Position encoding: Since Transformer itself does not have the ability to capture sequence order, position encoding is needed to provide the position information of the word fragment in the original text. Position encoding usually uses a combination of sine and cosine functions to add position information to each word fragment embedding.

[0116] 4. Encoded output: After being processed by multiple Transformer layers, the model finally generates an encoded output sequence. This sequence is a high-level representation of the input text.

[0117] The attention mechanism allows the neural network to automatically assign weights to words in the source sentence that are related to the predicted object and block irrelevant tags. In the specific social platform text stance detection, the model also needs to pay more attention to the language expression that is important to the event and block some unimportant information. In the standard attention mechanism, the tweet vector is represented as a weighted sum of hidden states, so there is no direct interaction between the tweet vector and the specific event in such a setting. The Softmax normalized weights used by the standard attention mechanism indirectly make the relationship between words, but this is inconsistent with human intuition. Intuitively, humans will only pay attention to some words related to a specific event. For example, a word in the text may imply the object's opposing stance, but this word is irrelevant to other events. The weight score of the word should be independent, that is, the relevance of each word to the event should be considered independently.

[0118] The Gated Attention Mechanism is a mechanism used in natural language processing tasks to enhance the model's ability to focus on specific parts of the text. In order to make stance detection more obvious, in this embodiment, the gating structure is used to expand the attention of the current word to a more fine-grained semantic level. The purpose is to allow the model to pay more attention to the part of the text related to a given object, thereby improving the accuracy of stance detection.

[0119] In order to independently consider the relevance of each word object, in this embodiment, by introducing the interaction between the hidden state and the object vector representation, the gating structure is used to expand the attention of the current word to a finer-grained semantic level. The hidden state of the tweet for a specific event is represented as:

[0120] b i =a i ⊙h i

[0121] The attention weight ai is used to determine the importance of hi to a specific event. This is calculated through a gating structure:

[0122] a i =σ(e i )

[0123] In the formula, ei is calculated using multi-layer perceptron attention, transforming the input through a multi-layer perceptron with a single hidden layer:

[0124] e i =V T (tanh(W h h i +W j j))

[0125] After obtaining the final sentence representation b, it is first transformed through a multi-layer perceptron, and the dimension of the multi-layer perceptron output is the number of possible stance categories. The output is sent to the Softmax layer and converted into a probability distribution. Since all parts of the model are differentiable, I can use standard back propagation to train in an end-to-end manner, and use multi-category cross entropy as the loss function, which can be expressed as follows:

[0126]

[0127] For the stance description in the text of the social platform, the topic content can be further combined. While performing stance detection, the topic description is also used to assist stance detection.

[0128] For the latent vectors output by Transformer of the original tweet sequence, take the average as the sentiment representation vector:

[0129]

[0130] Send r to the multi-layer perceptron for transformation, the output dimension is 1, that is, the predicted topic score u., and the mean square error is used as the loss function:

[0131]

[0132] The loss of combining the main task of stance detection and the auxiliary task of topic prediction can be expressed as:

[0133] L=λL main +(1-λ)L aux

[0134] After obtaining the emotional tendency of individual users, statistical methods can be used to aggregate group topics. Specifically, Fig.14 As shown, the frequency or weight of each sentiment label in the group can be calculated, and then the group can be divided into different stance categories based on these statistical information.

[0135] S105, taking the community as the perception object, constructing a cognitive confrontation perception picture according to the cognitive confrontation process of each community in the full time dimension of the specified event, wherein the full time dimension is from the beginning to the end of the event.

[0136] The community is the core of cognitive confrontation in cyberspace. In a sense, cognitive confrontation can be understood as the competition, integration and extinction between communities with different positions and themes. At present, there are two main types of popular cognitive domain perception perspectives. One is to use public opinion as the entry point to study the cognitive domain perception means of the development and evolution of online public opinion, but this obviously ignores the subjective status of people as cognitive confrontation; the other is to use a single point of entry to study the social behavior and position changes of individual users, but this is easily affected by individual variability and uncertainty, and obviously ignores the complex social network environment. These two methods have obvious shortcomings and it is difficult to correctly perceive the occurrence and changes of complete cognitive confrontation.

[0137] In the embodiment of this case, the community is used as the basic unit of cognitive confrontation research. By studying the attribute changes within the community, the attribute changes between communities and the global attribute changes, the global cognitive confrontation offensive is perceived.

[0138] Specifically, a multi-level cognitive perception model can be constructed based on the community cognitive confrontation process. The multi-level cognitive perception model perceives the dynamic evolution of public opinion positions on designated events within and between communities from the intra-group level, inter-group level and global level. The community structure and changes in community influence are used to measure the evolution of public opinion positions on designated events within the community, the number of inter-group edges and the flow of community personnel are used to measure the evolution of public opinion positions on designated events between communities, and the changes in the total number of communities, the number of nodes in each community and the dominant community are used to measure the evolution of public opinion positions on designated events in the global environment of the perceived social network.

[0139] Taking the community as the basic research object of cognitive confrontation is of revolutionary significance. In a specific social network, the positions and opinions of individuals are changeable, and the overall public opinion is abstract and non-specific. However, the existence and positions of the community are relatively stable. The emergence and disappearance of the community is a continuous dynamic process, which is observable and measurable. Taking the community as the basic unit gives a basic foothold to cognitive domain perception, so that cognitive domain perception is implemented from an abstract generalized concept to a measurable specific practice. In the process of cognitive domain perception, in the implementation examples of this case, attention is paid to the interactions and influences between communities, and how these interactions shape and change the cognitive structure and position of the community. Through in-depth analysis of the community and macroscopic perception of the cognitive domain, a more comprehensive and three-dimensional picture of cognitive confrontation in cyberspace is constructed.

[0140] like Fig.15As shown, first, the intra-group level focuses on the dynamics within a single community. At this level, the closeness and consistency between community members are evaluated by quantitatively analyzing the community's aggregation degree. The degree of aggregation reflects the degree of consensus among community members on positions and topics, as well as the cohesion and stability within the community. In addition, the changes in the community's positions and topics over time are analyzed to identify the evolution and differentiation of views within the community. By tracking the changes in the core members and marginal personnel of the community, we can further understand the power structure and influence distribution within the community, and thus perceive the competition and conflict within the community.

[0141] Secondly, the inter-group level focuses on the interaction and competition between different communities. At this level, the frequency and intensity of interaction between communities are evaluated by analyzing the establishment and disappearance of connections between communities. The establishment and breaking of connections reflect the formation and disintegration of relationships between communities and are a direct manifestation of competition and cooperation between communities. At the same time, attention is paid to the flow and destination of community members, and the patterns and trends of community members moving from one community to another are analyzed. In addition, the characteristic distance of communities is also an important indicator, which reflects the differences and similarities in communities' positions, topics, and behavior patterns, and helps to identify potential competition and cooperation opportunities between communities.

[0142] Finally, the global level focuses on the cognitive environment and public opinion atmosphere of the entire cyberspace. At this level, the cognitive diversity and dynamic changes of cyberspace are evaluated by monitoring the increase or decrease in the number of communities. The positions and views of the dominant communities reflect the mainstream cognition and values ​​of the entire cyberspace and are the key to understanding the overall trend of the cognitive domain. At the same time, pay attention to major changes in the direction of public opinion, analyze how the public opinion environment affects the positions and behaviors of the community, and how the community responds and adapts to these changes. In addition, by evaluating the impact of the public opinion environment on the community structure, it is possible to identify which communities disintegrate or reorganize under the pressure of public opinion, thereby perceiving the overall changes and evolution of the cognitive domain.

[0143] Through comprehensive analysis of these three levels, we can construct a comprehensive and three-dimensional cognitive domain perception model, which can not only reveal the complex relationships within and between communities, but also reflect the cognitive environment and public opinion trends of the entire cyberspace.

[0144] Intra-group competition starts from the level of a single community, and senses some key attribute changes within the community, so as to sense the direction of the community's cognitive changes and whether it is under attack. Some important indicators are selected to measure changes at the intra-group level.

[0145] 1. The total number of nodes N. The number of nodes is a direct reflection of the strength of the community. The increase in the number of nodes means the continuous expansion of the community, and the number of nodes returning to zero also means the demise of the community. By observing the changing trend of the total number of nodes N, we can perceive the change and growth trend of the community. If a community with a highly increasing number of nodes suddenly drops sharply, it is very likely that it has been attacked.

[0146] 2. Distribution and changes of core members. Social networks often exhibit the characteristics of scale-free networks, and there are certain dominant or influential nodes. The positions and opinions of these nodes often greatly affect the overall positions and opinions of the community. In order to measure these nodes, you can choose to extract the corresponding community as a subgraph and calculate the degree centrality of each node in the subgraph:

[0147]

[0148] Calculate the degree centrality of nodes in the community, sort and select the top 10% as core nodes. Analyze the changes in the community attributes of core nodes to detect whether they have been co-opted or infiltrated.

[0149] 3. Evaluation of community solidarity. Select a single community as a subgraph and analyze the changes in community solidarity. Select the clustering coefficient of the subgraph as an indicator.

[0150] The clustering coefficient of a single node is:

[0151]

[0152] The clustering coefficient of the entire community subgraph is the sum of the clustering coefficients of all nodes:

[0153]

[0154] By analyzing the changes in the clustering coefficient of the community, we can measure the changes in the organizational tightness within the community and perceive the changes in the tightness and solidarity of the members within the community.

[0155] At the same time, the modularity coefficient of the subgraph is selected as an indicator. The modularity coefficient reflects whether there is clustering in the network. Here it can be used to measure whether the community will split and generate new communities. The modularity coefficient is expressed as:

[0156]

[0157] By comprehensively analyzing the dynamic changes in the clustering coefficient and modularity of the community subgraph, we can perceive the structural changes of the community.

[0158] 4. The change in the average influence of the community is calculated by calculating the node centrality of each node in the community in the entire network, calculating and averaging the node centrality, and calculating the average centrality of the entire community.

[0159] The centrality of a single node in the entire community is:

[0160]

[0161] The average centrality of a community:

[0162]

[0163] The change in the average centrality of a community can help us see the changing trend of the community's influence in the entire network, and sense whether the community is gradually becoming dominant or gradually being isolated and marginalized.

[0164] 5. Changes in the community's stance and themes. Dynamically analyze and observe the changing trends of the community's stance and themes, perceive the changing process of the community's stance and themes, and combine the changes in the previous indicators and the overall changes in public opinion to clearly perceive whether the current community is being tempted, lured, or attacked and disintegrated.

[0165] Intergroup competition starts from the relationship between communities, perceives the changes in the relationship between communities, and thus perceives the offensive and defensive situation in the cognitive domain. Some important attributes of the interaction between communities are selected as measurement indicators to analyze and measure the relationship between groups.

[0166] 1. The number of connections between groups: by measuring the number of connections between a community and other communities, sorting by community divisions, and analyzing the changes in the number of connections. If a community suddenly establishes a large number of connections with this community, it can be sensed whether this community is attempting to infiltrate and assimilate; if a community with a large number of connections suddenly cuts off its connections, then it may be isolated or quarantined.

[0167] 2. The flow of community members. By analyzing the flow of community members in a time series, such as which community has the most members, we can determine the flow of community members. If there is a lot of flow between two members, it can be determined that there is infiltration in the flow of community members.

[0168] The global level is a key level for perceiving changes in the global environment. It analyzes the overall and global changes in network communities and network structures, and analyzes changes in overall cognitive trends and public opinion. The global level is analogous to the environmental conditions of the theory of natural selection. When the overall direction of public opinion changes, the number and overall distribution of communities will change greatly. Conversely, the overall changes in community distribution can be used to perceive whether there has been a drastic change in the direction of public opinion at the global level. Some important indicators are selected to perceive overall changes in the global situation.

[0169] 1. Changes in the total number of communities. Changes in the total number of communities can measure changes in the overall situation. When the total number of communities remains stable, it can be said that the overall situation is relatively stable. If the total number of communities changes significantly, it may mean that a major public opinion event has occurred;

[0170] 2. Changes in the number of nodes in each community. By measuring the dynamic changes in the number of nodes in each community in the space, we can measure the changes in the strength of each community and observe the distribution and changes of nodes in the community.

[0171] 3. Changes in the dominant communities on social platforms and their positions, views and narrative themes. By sorting the centrality of the previous communities, selecting the community with a higher centrality as the dominant community, and analyzing and observing the positions and themes of the dominant community, we can perceive the overall cognitive situation.

[0172] Furthermore, based on the above method, an embodiment of the present invention also provides a large model-driven social platform cognitive confrontation perception system based on dynamic collaborative detection, comprising: a data conversion module, a network coupling module, a topic detection module, a sentiment classification module and a confrontation perception module, wherein:

[0173] A data conversion module, used to construct a single social relationship feature network based on user behaviors in social network data, and to compound each single social relationship feature network to convert the social network data into a social relationship feature compound network;

[0174] The network coupling module is used to couple the composite networks of social relationship features in different time slices and obtain the community distribution on each time slice;

[0175] Theme detection module is used to detect the narrative theme of each community group based on the community distribution and the social text content in each community and using a large language model;

[0176] A sentiment classification module, which is used to obtain the community's stance on a specified event based on the narrative theme of each community group and using a pre-trained stance detection model, wherein the stance detection model includes a Transformer encoder for obtaining word embedding representations of both the specified event and the group narrative theme, and a text-gated attention mechanism for matching information between the specified event and the group narrative theme using word embeddings to determine the sentiment stance of the group narrative theme on the specified event;

[0177] The confrontation perception module is used to take the community as the perception object and construct a cognitive confrontation perception picture according to the cognitive confrontation process of each community in the full time dimension of the specified event, where the full time dimension is from the beginning to the end of the event.

[0178] like Fig.16The architecture shown can be divided into data layer, network layer, semantic layer, cognitive layer and front-end UI layer according to the level of data processing and analysis object. The specific functions and structures of each layer are as follows:

[0179] The data layer includes basic database and data preprocessing modules, which are mainly used for data preprocessing and data storage, and is the foundation of the system;

[0180] The network layer mainly includes the network construction subsystem and the dynamic collaboration detection subsystem, which are mainly used to establish social networks and perform dynamic collaboration detection, abstract social platform data into network structures and divide communities, and provide community structures of each time slice for downstream tasks;

[0181] The semantic layer mainly includes the group narrative analysis subsystem and the group stance detection subsystem. It mainly combines the community structure distribution and user text information, sinks to the semantic level to perform group narrative theme analysis and stance detection, and provides basic semantic information for subsequent cognitive perception models.

[0182] The cognitive layer mainly includes the cognitive confrontation perception subsystem, which combines the data information of all layers, perceives cognitive confrontation from the perspective of dynamic communities, and displays the perception data in a visual way to the front-end UI layer;

[0183] The front-end UI layer is mainly used to receive data transmitted by the cognitive layer, display the overall effect of the system, interact with users, and provide operational pages.

[0184] The front end imports the data into the database. At the network layer, the network construction subsystem constructs the initial data set into different relationship feature networks according to different relationships, and then composites the relationship feature networks; the dynamic collaboration detection subsystem couples different time slices and performs dynamic collaboration detection based on the composite graph. At the semantic layer, large model technology is used to analyze group narratives, and the group narrative results are integrated into the group stance detection subsystem to transmit stance and narrative theme data to the database. On the basis of the full-time dimension where the cognitive confrontation perception subsystem is located, the community is taken as the basic object of perception, and the process of cognitive confrontation is analyzed from the intra-group level, inter-group level, and global level, and a cognitive confrontation analysis diagram from the beginning to the end of the event is constructed. Finally, the calculation results are displayed on the front-end UI layer.

[0185] In order to verify the effectiveness of this solution, the following is a further explanation of the performance of this solution combined with experimental data:

[0186] The B / S structure is used and deployed on a cloud server. In response to the "Huinong" incident, the response speed, concurrency, compatibility and other aspects were tested. The following are the test results:

[0187] The response speed test results are as follows:

[0188] Table 1 Response test results

[0189]

[0190] The concurrent test results are as follows:

[0191] Table 2 Concurrency test

[0192]

[0193] The compatibility test results are as follows:

[0194] Table 3 Compatibility test results

[0195]

[0196]

[0197] After the above test verification, it can be proved that the solution in this case is credible in terms of response speed, concurrency and compatibility, and can meet the requirements of cyberspace purification and public opinion supervision.

[0198] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present invention.

[0199] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0200] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0201] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0202] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A large model-driven social platform cognitive adversarial perception method based on dynamic collaborative detection, characterized in that: Include: Constructing a single social relationship feature network based on user behaviors in social network data, and compounding each single social relationship feature network to transform the social network data into a social relationship feature compound network; Couple the composite networks of social relationship features in different time slices and obtain the community distribution on each time slice; Based on the distribution of communities and the content of social texts in each community, the narrative themes of each community group are detected using a large language model; Based on the narrative themes of each community group and using a pre-trained stance detection model, the stance detection model includes a Transformer encoder for obtaining word embedding representations of both the designated event and the group narrative theme, and a text-gated attention mechanism for matching the designated event and the group narrative theme using word embeddings to determine the emotional stance of the group narrative theme on the designated event; Taking the community as the perception object, a cognitive confrontation perception picture is constructed according to the cognitive confrontation process of each community in the full time dimension of a specified event, and the full time dimension is from the beginning to the end of the event.

2. The large model-driven social platform cognitive confrontation perception method based on dynamic collaborative detection according to claim 1 is characterized in that: Based on the user behavior in the social network data, a single social relationship feature network is constructed, and each single social relationship feature network is compounded, including: Taking individual user accounts in social platforms as the research objects, we construct the forwarding relationship feature network, the attention relationship feature network and the comment relationship feature network based on the user behavior of each individual user account; According to the distribution weight of each single social relationship feature network in the social network, the forwarding relationship feature network, the attention relationship feature network and the comment relationship feature network are merged into an undirected authorized network, and the undirected authorized network is used as the social relationship feature composite network.

3. The large model-driven social platform cognitive confrontation perception method based on dynamic collaborative detection according to claim 1 is characterized in that: The social relationship feature composite networks in different time slices are coupled and the community distribution on each time slice is obtained, including: The characteristic network structure of social relations at each time point is used as a time slice, and the network data at the corresponding time point is coupled by continuous time slices to construct a multi-dimensional time composite graph, which is used to describe the structural characteristics of the community at different time points and the changes in the relationships between community members; For multi-dimensional time composite graphs, the Leiden algorithm is used to separate nodes into disjoint communities, and nodes are moved recursively. The nodes and communities are merged by maximizing the quality function of each community to obtain the aggregated community distribution.

4. The large model-driven social platform cognitive confrontation perception method based on dynamic collaborative detection according to claim 3 is characterized in that: The Leiden algorithm is used to recursively merge and aggregate communities, including: Initialize a queue using all nodes; Remove the first node from the front of the queue and determine whether moving the node to the new community can increase the community quality function. If the community quality function increases, move the node to the new community and add neighbor nodes that do not belong to the new community and are not in the queue to the tail of the queue. Go back to the step where you removed the node from the front of the queue and repeat moving the removed node to the new community until the queue is empty.

5. The large model-driven social platform cognitive confrontation perception method based on dynamic collaborative detection according to claim 1 is characterized in that: Based on the distribution of communities and the content of social texts in each community, the narrative themes of each community group are detected using a large language model, including: In each time slice, the language type of social texts in the community is detected, and the community is divided into two levels based on language type; A large language model is used to perform topic analysis on the social text in each secondary community, and the topic analysis results of each secondary community are merged to obtain the narrative body of each overall community in each time slice. The large language model is TopicGPT, which uses LLMs to mine potential topics in a text collection through prompts.

6. The large model-driven social platform cognitive confrontation perception method based on dynamic collaborative detection according to claim 1 is characterized in that: Based on the narrative themes of each community group and using the pre-trained stance detection model, the community's stance on a given event is obtained, including: The word fragment model is used to decompose social text and specified events into a series of word fragments. Each word fragment is mapped to a word space vector using the Transformer encoder, and positional encoding is used to add position information to each word fragment to obtain the word embedding vector representation of the corresponding encoding sequence output; Use a gating structure to expand the attention of each word fragment to a fine-grained semantic level, use the interaction between hidden states and word vector representations to obtain sentence representations related to specified events in social texts, and use a multi-layer perceptron to obtain the number of emotion label categories and related probability distributions for specified events. Statistical methods are used to aggregate community group narrative themes, and the community group's stance on a given event is determined based on the frequency or weight of each emotional tag in the community group.

7. The large model-driven social platform cognitive confrontation perception method based on dynamic collaborative detection according to claim 1 is characterized in that: According to the cognitive confrontation process of each community in the full time dimension of the specified event, the cognitive confrontation perception picture is constructed, including: A multi-level cognitive perception model is constructed based on the community cognitive confrontation process. The multi-level cognitive perception model perceives the dynamic evolution of public opinion positions on designated events within and between communities from the intra-group level, inter-group level and global level. The evolution of public opinion positions on designated events within the community is measured by changes in community structure and community influence, the evolution of public opinion positions on designated events between communities is measured by the number of inter-group edges and the flow of community personnel, and the evolution of public opinion positions on designated events in the global environment of the social network is measured by changes in the total number of communities, the number of nodes in each community and the dominant community.

8. A large model-driven social platform cognitive confrontation perception system based on dynamic collaborative detection, characterized in that: It includes: data conversion module, network coupling module, topic detection module, sentiment classification module and adversarial perception module, among which, A data conversion module, used to construct a single social relationship feature network based on user behaviors in social network data, and to compound each single social relationship feature network to convert the social network data into a social relationship feature compound network; The network coupling module is used to couple the composite networks of social relationship features in different time slices and obtain the community distribution on each time slice; Theme detection module is used to detect the narrative theme of each community group based on the community distribution and the social text content in each community and using a large language model; A sentiment classification module, which is used to obtain the community's stance on a specified event based on the narrative theme of each community group and using a pre-trained stance detection model, wherein the stance detection model includes a Transformer encoder for obtaining word embedding representations of both the specified event and the group narrative theme, and a text-gated attention mechanism for matching information between the specified event and the group narrative theme using word embeddings to determine the sentiment stance of the group narrative theme on the specified event; The confrontation perception module is used to take the community as the perception object and construct a cognitive confrontation perception picture according to the cognitive confrontation process of each community in the full time dimension of the specified event, where the full time dimension is from the beginning to the end of the event.

9. An electronic device, characterized in that: include: at least one processor, and a memory coupled to the at least one processor; The memory stores a computer program, and the computer program can be executed by the at least one processor to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 7 can be implemented.

Images