A Social Robot Intent Recognition System and Method Based on Heterogeneous Graph Analysis

Abstract

This invention provides a social robot intent recognition system and method based on heterogeneous graph analysis. Based on a heterogeneous graph structure, it uniformly models multiple types of nodes and multiple relational edges, systematically capturing complex multidimensional behavioral relationships within social platforms. It integrates technologies such as large language models and knowledge graphs to support refined identification and classification of multiple behavioral patterns. Simultaneously, by combining account behavior aggregation and sentiment polarity analysis, it achieves hierarchical assessment of different organizational risks, providing a basis for regulatory and governance decisions. It integrates automatic collection, preprocessing, and heterogeneous graph construction processes for multi-source heterogeneous data, minimizing manual intervention and repetitive work, improving overall data processing efficiency and response speed, and meeting the real-time analysis needs of complex and dynamic social environments. It supports transparent analysis of robot control stances and behavioral strategies, greatly enhancing the system's interpretability, user trust, and the enforceability of policy regulations.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to a social robot intent recognition system and method based on heterogeneous graph analysis. Background Technology

[0002] With the continuous advancement of artificial intelligence technology, especially the increasing maturity of natural language processing and social network analysis methods, the identification and intent analysis of abnormal accounts (such as social bots) on social platforms has gradually become an important research direction in fields such as cybersecurity, information dissemination control, and public opinion analysis. Existing technologies mainly include rule-based matching identification methods, machine learning methods based on user behavior modeling, and text analysis methods relying on language models. However, these methods generally suffer from low identification accuracy, insufficient generalization ability, and lack of interpretability in practical applications, making it difficult to effectively model the complex relationships between various types of objects on social platforms and limiting their application effectiveness in complex social environments.

[0003] Specifically, current technology has the following limitations:

[0004] (1) Lack of rule or behavior feature recognition methods that combine semantic understanding and multimodal fusion

[0005] Traditional social bot identification methods often rely on manually constructed behavioral rules (such as posting frequency, posting time distribution, and follower-follower relationships) or superficial statistical features based on user behavior data. These methods neglect the semantic information and underlying intentions contained in the content posted by accounts, making it difficult to identify the complex behavioral patterns of bot accounts in areas such as stance manipulation, emotional bias intervention, and topic implantation. Furthermore, these rules are poorly adaptable to platform differences and diverse expressions, making them easily circumvented by malicious accounts with adversarial camouflage capabilities.

[0006] (2) Language models lack the ability to model structured relationships in text recognition.

[0007] In recent years, large language models such as BERT and GPT have achieved significant results in semantic understanding and content generation. Some studies have applied them to social content analysis to identify bot accounts and their information manipulation intentions. However, language models mainly focus on processing local semantic information and lack the ability to model the complex structural relationships between diverse and heterogeneous elements such as users, posts, related entities, and stances. Their analysis scope is mostly limited to the text fragments themselves, ignoring the information propagation structure, entity referencing networks, and stance association networks in social networks, thus resulting in analysis results that lack contextual logical support and a global perspective.

[0008] (3) Lack of a unified framework for the integration and systematic analysis of heterogeneous relationships

[0009] Existing research struggles to achieve unified modeling of heterogeneous nodes and their relationships across multiple types, including accounts, posts, entities, and stances, and lacks a comprehensive analytical framework with structural expressive capabilities. When faced with the task of recognizing the intent of social robots, traditional methods often focus on a single dimension (such as behavior or text), failing to capture their deep, multi-layered, multi-perspective, and collaborative characteristics.

[0010] Therefore, there is a need for an intelligent recognition method that can integrate multimodal heterogeneous information and possess graph structure representation and semantic reasoning capabilities to improve the accuracy, efficiency and interpretability of recognition, and meet the technical requirements for social robot intent recognition in complex social environments. Summary of the Invention

[0011] This invention addresses the shortcomings in accuracy, efficiency, and interpretability in social bot identification by providing a social bot intent recognition system and method based on heterogeneous graph analysis. Based on a heterogeneous graph structure, it uniformly models multiple types of nodes (accounts, posts, entities, stances) and multiple relational edges, systematically capturing complex multidimensional behavioral relationships within social platforms. This effectively improves the model's ability to express and discriminate social bot manipulation behaviors, significantly enhancing the identification of abnormal accounts and their potential organizational intentions. Furthermore, it integrates large language models and knowledge graph technologies to support refined identification and classification of multiple behavioral patterns. By combining account behavior aggregation and sentiment polarity analysis, it enables hierarchical assessment of risks across different organizations, providing a basis for regulatory and governance decisions. It integrates automated collection, preprocessing, and heterogeneous graph construction processes for multi-source heterogeneous data, minimizing manual intervention and repetitive work, improving overall data processing efficiency and response speed, and meeting the real-time analysis needs of complex and dynamic social environments. Through graph visualization and behavior path tracing functions, it intuitively presents the semantic relationships and information propagation links between accounts, supporting transparent analysis of bot control positions and behavioral strategies, greatly enhancing the system's explainability, user trust, and the enforceability of policy regulations.

[0012] This invention provides a social robot intent recognition system based on heterogeneous graph analysis, comprising a heterogeneous social behavior graph construction and analysis module, an account-post-entity object-position bipartite graph network construction module, and a social robot organization intent recognition module connected in sequence.

[0013] The heterogeneous social behavior graph construction and analysis module captures the complex behavioral relationships and network structure characteristics of social robots on the platform by constructing heterogeneous social behavior graphs, embedding social features of account nodes and calculating dynamic edge weights, embedding propagation link features of account nodes, and analyzing account communities, thereby revealing collaboration patterns and obtaining highly collaborative account subgroups.

[0014] The Account-Post-Entity-Stance Bipartite Graph Network Construction Module extracts the target semantic subgraph structure of highly collaborative account subgroups, constructs the Account-Post Behavior Graph, Post-Entity Semantic Graph, Post-Stance Semantic Graph, and Account-Entity-Stance Semantic Graph, and generates the Account-Post-Entity-Stance Bipartite Graph Network. By constructing semantic subgraphs, it enables a fine-grained understanding of the manipulated objects and stance intentions, and characterizes the stance tendencies of each account.

[0015] The social robot's organizational intent recognition module uses a large language model and a heterogeneous graph attention network model to perform account goal analysis, sentiment stance consistency analysis, collaborative behavior analysis, and social network structure role analysis on accounts in the account-post-entity-position bipartite graph network to obtain multimodal features. It then performs account pattern recognition, visualization, and path tracing on the multimodal features to identify and predict the organizational behavior patterns and intents of accounts, and outputs the account's behavior pattern labels.

[0016] The social robot intent recognition system based on heterogeneous graph analysis described in this invention, as a preferred embodiment, includes a social robot intent recognition module comprising a large language model, a heterogeneous graph attention network model, a time series model, a graph attention network model, and a heterogeneous graph neural network model;

[0017] The social robot's organizational intent recognition module extracts the relationship structure diagram of account-post-entity object from the bipartite graph network of account-post-entity object. ,exist The semantic embedding vectors of the post nodes are obtained from the large language model to obtain semantic information.

[0018] Will Input a heterogeneous graph attention network model and output a label indicating whether an account is attacking or following a target. At the same time, the heterogeneous graph attention network model propagates semantic information from post nodes to account and entity nodes.

[0019] The time series model identifies behavioral features based on the time density vector sequence of accounts and outputs behavioral feature vectors; based on the similarity of behavioral feature vectors, highly synchronized subgroups are identified, thereby obtaining a group of accounts with highly coordinated actions.

[0020] Graph attention network models identify the influence of each node in a subgraph of highly collaborative action account groups;

[0021] Heterogeneous graph neural networks learn semantic interaction and collaborative behavior maps between accounts in multimodal features through interlayer attention mechanisms, and output behavioral pattern labels of accounts. The multimodal features are feature vectors that integrate content semantic vectors, sentiment consistency vectors, temporal behavior pattern vectors, and social structure embedding vectors.

[0022] The social robot intent recognition system based on heterogeneous graph analysis described in this invention, as a preferred embodiment, includes the following steps:

[0023] S1. Construction and Analysis of Heterogeneous Social Behavior Graphs: The module for constructing and analyzing heterogeneous social behavior graphs defines node types and relationship types, maps structured data to nodes and edges, and then calculates the social feature embedding vector and dynamic edge weights of accounts through social feature embedding and dynamic edge weights. Then, it obtains the propagation link feature embedding vector of accounts through the propagation link feature embedding of account nodes. Based on the social feature embedding vector, dynamic edge weights, and propagation link feature embedding vector, community discovery is performed to obtain highly collaborative account subgroups.

[0024] The S2 module, which constructs a bipartite graph network for accounts, extracts relevant nodes and relationships from highly collaborative account subgroups. It then extracts the structure of the target semantic subgraph and establishes edge structures to obtain an account-post behavior graph. This graph identifies named entities and confidence levels in posts and establishes edge weights to obtain a post-entity semantic graph. Finally, it analyzes post stance scores to establish stance edges, resulting in a post-stance semantic graph. A ternary relationship between account, entity, and stance is constructed, and weights are calculated to obtain an account-entity-stance semantic graph. These graphs are then fused to form an account-post-entity-object-stance bipartite graph network. This network is used to analyze the attack targets, stance biases, and content patterns of social bots in a specific semantic space, supporting intent recognition tasks.

[0025] S3, the social robot organization intent recognition module, utilizes a large language model and heterogeneous graph attention network to identify account targets, recognizing groups of accounts that frequently speak around the target entity. Based on these identified groups, it performs sentiment consistency analysis, analyzing the sentiment polarity of the group accounts, aggregating the group's sentiment scores towards the entity, and determining whether the group's stance and sentiment towards the target entity are consistent. Then, for groups with consistent stance and sentiment towards the target entity, it performs collaborative behavior analysis, calculating the synchronization degree between accounts, identifying highly synchronized subgroups as a basis for judging organized attack groups, and outputting highly coordinated action account groups. Next, it performs social network structure role analysis on the highly coordinated action account groups, identifying the roles of the accounts; for accounts with identified roles, it performs account pattern recognition to determine the intent of the account patterns; finally, it maps the intent recognition results to a graph for visualization and path tracing, outputting a user-friendly graphical view and report for regulators, along with account behavior pattern tags.

[0026] The social robot intent recognition system based on heterogeneous graph analysis described in this invention, in a preferred embodiment, includes the following steps in step S1:

[0027] S11. Construction of Heterogeneous Social Behavior Graph: Based on the collected account information and post information, the Heterogeneous Social Behavior Graph Construction and Analysis Module uses a large language model to encode the post text and defines the node type and relation type of the heterogeneous social behavior graph, resulting in a heterogeneous graph G, which includes nodes V and edges R.

[0028] S12. Account Node Social Feature Embedding and Edge Weight Calculation: The heterogeneous social behavior graph construction and analysis module uses a heterogeneous graph neural network for feature embedding, mapping the information of each account node in the heterogeneous graph G into a low-dimensional social feature embedding vector. Dynamic edge weights are calculated using a dynamic edge weight calculation mechanism based on content similarity, temporal proximity, and interaction frequency between accounts. These dynamic edge weights represent the collaborative relationship weights between accounts. Based on the social feature embedding vectors and the dynamic edge weights, the account social heterogeneous graph is obtained. Social heterogeneity graph of accounts This includes account node embedding features and collaborative relationships;

[0029] S13. Embedding of Account Propagation Link Features: The heterogeneous social behavior graph construction and analysis module first generates node sequences based on meta-path-guided random walks, and then uses a skip-gram model to train, learn, and optimize the semantic vector representations of nodes in the node sequences, emphasizing semantic co-occurrence in the propagation path. relation The Skip-gram collinear embedding vectors are obtained as the propagation link feature representation of the account, and then the social heterogeneity graph of the account is fused. The Skip-gram collinear embedding vectors yield a multidimensional heterogeneous graph that simultaneously includes social structure information and propagation information. ;

[0030] S14. Using a graph community detection algorithm based on multidimensional heterogeneous graphs. Calculate and maximize modularity Find the group of nodes that collaborate most closely, identify the community structure where multiple accounts cooperate in publishing and forwarding content from the same source, and obtain a subgroup of highly collaborative accounts.

[0031] In the social robot intent recognition system based on heterogeneous graph analysis described in this invention, as a preferred embodiment, in step S11, the node types include accounts, posts, and comments, the relationship types include posting / liking relationships, quoting / forwarding relationships, and commenting / liking relationships, and the edges in the relationship types all include timestamp information;

[0032] In step 12, the social feature embedding vector represents the identity representation of the current node after aggregating all semantic and structural information in the heterogeneous graph G; the content similarity is cosine similarity, the temporal proximity is an exponential decay function, and the interaction frequency between accounts includes forwarding, commenting, and liking;

[0033] In step S13, the jump character model learns vectors through continuous iterative training, which make the co-occurring nodes closer together, thus obtaining the co-occurrence vector representation of each node;

[0034] Skip-gram collinear embedding vector is a concatenated vector representing social feature embedding vector and co-occurrence vector;

[0035] In step S14, the graph community discovery algorithm is the Louvain algorithm, and the modularity Q is a function of the edge weight between the two accounts, the node weighted degree, the total weight of all edges, and the expected edge weight under the random network model. The higher the value, the more closely connected the communities are internally, and the more sparsely connected the communities are between each other.

[0036] The social robot intent recognition system based on heterogeneous graph analysis described in this invention, in a preferred embodiment, includes the following steps in step S2:

[0037] S21, the Account-Post-Entity-Stance Bipartite Graph Network Construction Module selects specific social bot communities from highly collaborative account subgroups and extracts relevant nodes and edges within the subgraph to obtain the initial semantic analysis subgraph. ;

[0038] S22, From the initial semantic analysis subgraph Extract the node set and edges, and assign weights to each edge to obtain the account-post behavior graph. The account-post behavior graph reveals the frequency and pattern of the account's language operations, forming the behavioral background graph for intent analysis.

[0039] S23. Use a large language model to identify named entities in the post and extract the entity set. Perform disambiguation on all entities to generate the final entity nodes. Use the confidence output by the large language model to construct the edges from the post nodes to the entity nodes to obtain the post-entity semantic graph.

[0040] S24. Use large language models to analyze the sentiment towards the target entity in the posts. The results of the stance judgment on the post are obtained, and the edges and weights are constructed based on the stance judgment results to obtain the post-stance semantic graph;

[0041] S25. Integrate the post-entity semantic graph and the post-position semantic graph into the account dimension, extract effective triple paths to construct the account-entity-position semantic graph, and form a semantic intent graph.

[0042] S26. By integrating the account-post behavior graph, post-entity semantic graph, post-position semantic graph, and account-entity-position semantic graph, a bipartite graph network of account-post-entity-object-position is obtained. By constructing the bipartite graph network of account-post-entity-object-position, the posts published by the account, the corresponding position status of the posts, and the target entities mentioned are clarified. The position relationship between the account and the entity is obtained through the position corresponding to the account post and the mentioned entity.

[0043] In the social robot intent recognition system based on heterogeneous graph analysis described in this invention, as a preferred embodiment, in step S21, a specific social robot community is a highly collaborative subgraph with a tightly structured propagation.

[0044] In step S22, the weight of each edge is a function of the account's influence and the similarity between the post and the account's historical content;

[0045] In step S23, the large language model is ChatGPT or DeepSeek, and the named entities are the targets of post attacks / attention, including people's names, organization names, country / region, event names, and topics; disambiguation is performed using the large language model, business rule base conversion, or embedding vector similarity calculation method; confidence is used to replace the weights representing mention strength or relevance.

[0046] In step S24, emotional stance Including support, opposition, and neutrality, the post-position semantic graph depicts the stance and emotional inclination of each text.

[0047] In step S25, all post nodes under the account are traversed. If a post node has an edge with both an entity and a position, it is determined that the post under the account expresses this position of the entity, forming a triplet path:

[0048] The weighting function reflecting the intensity of the emotions expressed by an account towards an entity is based on all accounts that simultaneously mention the entity. And express a stance It is a function of the set of posts, the behavioral intensity of accounts and posts, the confidence of entities mentioned in posts, and the probability score of the stance expressed in posts. The behavioral intensity of posts includes frequency and style similarity.

[0049] In the social robot intent recognition system based on heterogeneous graph analysis described in this invention, as a preferred method, in step S3, when identifying the account target, the heterogeneous graph attention network model aggregates features from the node's neighbors through weighted average aggregation, updates the representation of the current node, and assigns different weights to each post according to the account semantic representation. Based on the account semantic representation and the semantic popularity of entity objects, an indirect semantic path between the account and the entity is constructed to obtain the content semantic vector, thereby understanding the intensity of the account's attention / attack on the entity; the account semantic representation is the semantic focus of the account, and the semantic popularity of entity objects is the degree of semantic focus of the entity being discussed;

[0050] During sentiment consistency analysis, a large language model is used to perform sentiment analysis on the target post text to obtain a sentiment polarity score. The sentiment polarity score ranges from -1 to 1, including positive, negative, and neutral.

[0051] When entity When mentioned in multiple posts, the average sentiment score of the entity is obtained by summing the sentiment scores of the posts.

[0052] The sentiment average of all posts by an account that mention the entity reveals the account's sentiment towards that entity:

[0053] Finally, an entity was established for a group of accounts, U. Consistency analysis:

[0054] .

[0055] In the social robot intent recognition system based on heterogeneous graph analysis described in this invention, as a preferred embodiment, in step S3, during collaborative behavior analysis, a multidimensional heterogeneous graph is used. For each account u in group U, obtain all post timestamps and divide the time window according to time unit to obtain a time density vector sequence:

[0056] ;

[0057] in, Let be the number of posts in the k-th time window;

[0058] time density vector sequence The data is input into a time series model to obtain the behavioral feature vector of the account and a highly synchronized subgroup. The behavioral features include the account's posting rhythm, periodicity, and whether it posts day and night.

[0059] Based on the number of members in highly synchronized subgroups, the probability of consistency of sentiment stance, and the magnitude of edge weights, determine whether each subgroup is involved in organized attacks, and output a group of accounts with highly coordinated actions.

[0060] The social robot intent recognition system and method based on heterogeneous graph analysis described in this invention, as a preferred embodiment, uses a graph attention network model to identify the influence of each node in a subgraph of highly collaborative action account groups during social network structure role analysis to obtain the collaboration strength. :

[0061] ;

[0062] in, For account The initial embedding is a concatenation of the temporal feature vector, content embedding vector, and sentiment consistency vector of the account node; TM is a trainable linear transformation matrix used for dimensionality reduction or transformation of the embedding space, which is obtained by training a graph attention network model. For a trainable attention weight vector, and The attention score is obtained by multiplying the results, which is derived from training a graph attention network model. This indicates that two vectors have been concatenated; LeakyReLU is the activation function. For nodes Normalize all neighbor nodes j so that ;

[0063] node The total attention weight is: ;

[0064] Divide the number of edges of a node by the total number of nodes to get the number of nodes. Degree centrality ;

[0065] Get Node First attack timestamp ;

[0066] Based on degree centrality Attention value First attack timestamp Defining node roles:

[0067] Determine the account when the following conditions are met. The node role is the commander node:

[0068] ;

[0069] in, For degree centrality threshold, The maximum degree centrality of all accounts in the current account group. Attention weight threshold ≥0.7, This represents the sum of the maximum attention scores of all accounts in the group. Allowing for advance planning, The time of the earliest attack post in the group;

[0070] Determine the account when the following conditions are met. The node role is that of a responder:

[0071] ;

[0072] in, For account The number of incoming edges, This is the threshold coefficient for the number of incoming edges. For account The number of outgoing edges indicates the degree to which it actively influences others. This is the threshold coefficient for the number of outgoing edges. For account Similarity of content with the leader's account in the group. This is the similarity threshold coefficient;

[0073] Determine the account when the following conditions are met. The node role is an external account:

[0074] ;

[0075] in, For a lower attention weight threshold, ≤1 / 3 , The threshold corresponding to excessively low content similarity. The average time to the first post for the account group. The time delay threshold is used; the account's behavioral pattern tags include: concentrated attack, public opinion diversion, identity spoofing, and emotional manipulation.

[0076] This invention relates to the technical fields of artificial intelligence and natural language processing, graph neural networks and social network analysis.

[0077] (1) This invention provides a social robot intent recognition method based on heterogeneous graph analysis, which integrates heterogeneous information of multiple types such as accounts, posts, entities, and stances to construct a heterogeneous graph with structural representation capabilities, and introduces a graph neural network model for multidimensional semantic reasoning and intent recognition. This invention significantly improves the ability to identify abnormal accounts and the accuracy of interpreting behavioral patterns on social platforms, and enhances the ability to model potential manipulation paths in complex social contexts.

[0078] (2) This invention addresses the problem that traditional identification methods lack deep semantic understanding and structural modeling capabilities. Existing social robot identification methods mainly rely on surface behavioral features or rule templates, making it difficult to reveal their deep semantic logic in terms of stance manipulation, emotion guidance, and topic intervention. This invention constructs a heterogeneous graph that integrates accounts, posts, cited entities, and stance orientations, systematically expressing the potential paths and influence mechanisms of robot accounts in the information dissemination chain, significantly improving the semantic depth and structural explanatory power of identification.

[0079] (3) This invention overcomes the shortcomings of language models in structured information modeling and cross-entity semantic fusion. Although large language models have advantages in natural language understanding, their ability to express the structured semantics of multiple entities and relationships in social platforms is limited, making it difficult to capture the high-order associations between user behavior and semantic objects. This invention introduces a heterogeneous graph neural network architecture to integrate text semantics, user behavior, entity references, and stance orientation into a structured modeling framework, achieving cross-entity and cross-level semantic fusion and behavior pattern recognition, thus solving the problem of limited granularity in traditional language model analysis.

[0080] (4) This invention breaks through the bottleneck of single-dimensional modeling in social platform content analysis. Existing technologies usually focus on a single dimension, such as user behavior, text features, or social relationships, and fail to form a systematic expression of the interaction logic between multiple types of nodes, resulting in one-sided intent recognition results and contextual disconnect. This invention constructs a heterogeneous graph network structure that integrates account-post-entity-stance relationship, and combines layer cascading and message propagation mechanisms to achieve panoramic intent judgment of "who is saying what, to whom, and what stance is being expressed", effectively improving the comprehensiveness and accuracy of social behavior analysis.

[0081] (5) This invention improves the interpretability and cross-platform generalization ability of the social robot recognition system. Existing recognition schemes that rely on language models are mostly black-box mechanisms, lacking traceable path logic, and are difficult to meet the needs of regulatory compliance and policy response. This invention utilizes the explicit node relationships and semantic propagation paths in the graph structure to support the visualization and tracking analysis of the recognition results, thereby enhancing the interpretability of the recognition mechanism. At the same time, through the unified modeling framework of heterogeneous graphs, it has good structural universality and semantic transfer ability, which helps to realize cross-domain generalization applications in different platforms and different corpus environments.

[0082] The present invention has the following advantages:

[0083] (1) Improve the accuracy and intelligence of social robot intent recognition

[0084] Based on heterogeneous graph structure, a unified model of multiple types of nodes (accounts, posts, entities, stances) and multiple relational edges is used to systematically capture complex multidimensional behavioral relationships in social platforms. This effectively improves the model's ability to express and distinguish the manipulation behavior of social robots, thereby significantly enhancing the identification effect of abnormal accounts and their potential organizational intentions.

[0085] (2) Achieve multi-dimensional, fine-grained intent classification and risk assessment

[0086] By integrating technologies such as big data prediction models and knowledge graphs, it supports the refined identification and classification of various behavioral patterns, including concentrated attacks, public opinion manipulation, identity spoofing, and emotional manipulation. At the same time, by combining account behavior aggregation and sentiment polarity analysis, it enables hierarchical assessment of different organizational risks, providing a basis for decision-making in supervision and governance.

[0087] (3) Improve data processing and analysis efficiency

[0088] The system integrates automated collection, preprocessing, and heterogeneous graph construction processes for multi-source heterogeneous data, minimizing manual intervention and repetitive work, improving overall data processing efficiency and response speed, and meeting the real-time analysis needs in complex and dynamic social environments.

[0089] (4) Enhance the visualization and interpretation capabilities of the results.

[0090] Through graph visualization and behavior path tracing functions, the semantic relationships between accounts and information propagation links are presented intuitively, supporting transparent analysis of the robot's control stance and behavior strategies, which greatly improves the system's explainability, user trust, and the enforceability of policy supervision. Attached Figure Description

[0091] Figure 1 This is a flowchart of a social robot intent recognition method based on heterogeneous graph analysis;

[0092] Figure 2 This invention relates to a system and method for identifying the intent of social robots based on heterogeneous graph analysis, specifically focusing on robot community discovery based on heterogeneous social behavior graphs.

[0093] Figure 3 This is an example diagram of the bipartite network relationship of account-post-entity-position for a social robot intent recognition system and method based on heterogeneous graph analysis. Detailed Implementation

[0094] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example 1

[0095] A social robot intent recognition system and method based on heterogeneous graph analysis is provided. The social robot intent recognition system based on heterogeneous graph analysis includes a heterogeneous social behavior graph construction and analysis module, a "account-post-entity object-stance" bipartite graph network construction module, and a social robot organizational intent recognition module.

[0096] The heterogeneous social behavior graph construction and analysis module captures the complex behavioral relationships and network structure characteristics of social robots on the platform, revealing collaboration patterns.

[0097] The heterogeneous social behavior graph construction and analysis module collects and preprocesses multi-source social data: it captures account profiles, post content, reposts, comments, likes, time, etc., and outputs structured raw data (account, post, comment nodes and time information) after preprocessing.

[0098] Construct a heterogeneous social behavior graph: Define node types (User, Post, Comment) and relationship types (post, forward / quote, comment / like), and map structured data to nodes and edges;

[0099] Social feature embedding and dynamic edge weight calculation: Obtain the social feature embedding vector of the account and the collaborative relationship weight between accounts;

[0100] Propagation link feature embedding: Obtain the propagation link feature embedding vector of the account;

[0101] Community detection and identification: Based on social feature embedding, propagation link feature embedding and dynamic edge weights, highly collaborative account subgroups (such as suspected bot collaboration groups) are obtained through community discovery.

[0102] The “Account-Post-Entity-Stance” bipartite graph network construction module constructs semantic subgraphs to support fine-grained understanding of manipulated objects and stance intentions.

[0103] The "Account-Post-Entity-Stance" bipartite graph network construction module extracts the target semantic subgraph structure: from the highly collaborative account subgroups identified in the previous module, relevant User, Post, and Comment nodes and relationships are extracted.

[0104] Construct an "account-post" behavior graph: establish edges between accounts and their posts.

[0105] Construct a semantic entity graph of “post-entity”: identify named entities and confidence levels of posts, and establish edge weights.

[0106] Construct a semantic stance graph of "post-stance": analyze the stance score of posts (support / oppose / neutrality + probability) and establish stance edges.

[0107] Construct a ternary semantic graph of “account-entity-position”: construct the ternary relationship of (account, entity, position) and calculate the weight.

[0108] The above four types of diagrams are combined into a two-part diagram: "Account – Post – Entity – Stance".

[0109] The social robot's organizational intent recognition module identifies the account's organizational behavior patterns and intent based on a bipartite graph of "account-post-entity-position" and multimodal features.

[0110] The social robot's intent recognition module identifies account targets by frequently engaging with target entities.

[0111] Sentiment Consistency Analysis: Based on the account groups identified by account target identification, the sentiment polarity of the group accounts is analyzed, the sentiment scores of the account groups towards the entity are aggregated, and it is determined whether the group's stance and sentiment towards the target entity are consistent.

[0112] Collaborative Behavior Analysis: For groups of target entities with consistent stances and sentiments, calculate the synchronization degree of behavior between accounts, identify highly synchronized subgroups, use this as a basis for judging the group of organized attacks, and output the group of accounts with highly coordinated actions.

[0113] Social network structural role analysis: Identify the roles (leaders, responders, etc.) of accounts in highly collaborative action groups.

[0114] Account pattern recognition: For accounts with identified roles, determine the intent of the account pattern (targeted attack, public opinion guidance, identity disguise, emotional manipulation).

[0115] Visualization and path tracing: Maps intent recognition results to a graph, supports path tracing, and outputs user-friendly graphical views and reports for regulators.

[0116] The figure shows a flowchart of a social robot intent recognition method based on heterogeneous graph analysis, including the following steps:

[0117] Step S1: Construction and Analysis of Heterogeneous Social Behavior Maps

[0118] The aim is to capture the complex behavioral relationships and network structure characteristics of social bots on the platform, model structural features such as behavioral links and propagation paths, and reveal organizational-level propagation collaboration patterns, such as multi-account coordinated posting, forwarding of content from the same source, and attacking public opinion hotspots, to help identify "who is collaborating with whom". This mainly includes the construction of heterogeneous social behavior graphs, embedding of social features of account nodes and calculation of edge weights, embedding of propagation link features of account nodes, and analysis of account communities.

[0119] S11. Construction of Heterogeneous Social Behavior Graphs:

[0120] Based on the collected account and post information, after text encoding the posts using a large language model, the node types and relationship types of the heterogeneous social behavior graph are defined, where:

[0121] Node types include User, Post, and Comment.

[0122] ;

[0123] The relationship types include post / like relationships (User to Post, with edges including timestamp information), quote / forward relationships (Post to Post, with edges including timestamp information), and comment / like relationships (User to Comment, with edges including timestamp information).

[0124] Taking the account "Fourth Lane" as an example, the account "Fourth Lane" posted post p1 at 9:41 on June 16, 2025. The account "Road Traveler" commented on it, recorded as c1, at 10:20 on June 16, 2025.

[0125] The nodes that make up the structure:

[0126] 1) "The Fourth Lane", "Car Soul", "Road Traveler"

[0127] 2) p1

[0128] 3) c1

[0129] The edges formed:

[0130] 1) Release: ("Fourth Lane" → p1)

[0131] 2) Forward: (“Car Soul” → p1)

[0132] 3) Comment: (“Road Traveler” → c1)

[0133] The resulting heterogeneous graph G consists of nodes V and edges R.

[0134] ;

[0135] S12. Account Node Social Feature Embedding and Edge Weight Calculation:

[0136] First, a heterogeneous graph neural network (HGNN) is used for feature embedding, which maps the information of each account node in the graph (including the node's own attributes and the structural information of its neighboring nodes) into a low-dimensional vector representation (Embedding1).

[0137] The Heterogeneous Graph Neural Network (HGNN) needs to be trained. The model's input is a heterogeneous graph G, and its output is a pre-labeled group of social bots of different types. The core formula of HGNN is as follows:

[0138] ;

[0139] in, Let v be the representation of the node at level l. Let v be the set of neighbors of node v under relation r. A set of edge types, The normalization constant is , For relation-specific and self-loop weight matrices, It is a non-linear activation function. Therefore, This is represented as the "identity representation" of the current node after all semantic and structural information has been aggregated in the graph, and it serves as the input for the classification task in the HGNN.

[0140] Therefore, using the trained HGNN for feature embedding of each account node can accommodate all its information to a greater extent, thus enabling better identification of social bots from different groups.

[0141] Secondly, the edges between account nodes represent social relationships, and the edge weights represent the strength of those relationships. A dynamic edge weight calculation mechanism is proposed, and the calculation method is as follows:

[0142] ;

[0143] ;

[0144] ;

[0145] ;

[0146] in, For content similarity (such as cosine similarity); For time proximity (such as exponential decay function); The frequency of interaction between accounts (including forwarding, commenting, liking, etc.). , , For weight hyperparameters.

[0147] For example, the post by "Fourth Lane" was published on June 16, 2025 at 08:20; "Road Traveler" commented on it on June 16, 2025 at 10:20. The cosine similarity after vectorizing the post and comment is 0.84. The time difference between publication is 0.083 days (2 hours). The maximum tolerance time window is set to 7 days, the decay coefficient is customized to 3, and the time proximity is 0.96. If "Fourth Lane" and "Car Soul" have been forwarded 2 times, commented on 1 time, and liked 1 time in the past 7 days, the total number of interactions is 4. The maximum number of interactions between other users in the sample is 5, and the interaction frequency is 0.8. , , If the values ​​are 0.5, 0.3, and 0.2 respectively, then the two accounts have a collaborative relationship, with a weight of 0.94.

[0148] The final result is a heterogeneous social graph of accounts with embedded features and collaborative relationships. .

[0149] ;

[0150] S13. Account propagation link feature embedding:

[0151] Key propagation links are defined through meta-paths, for example:

[0152] .

[0153] This path pattern guides subsequent random walks, ensuring that the sampled sequences have specific semantic associations. Following the meta-path pattern, starting from a node, a random walk is performed along specified node and edge types to generate a node sequence. This node sequence resembles a word sequence in a sentence, where nodes "co-occur" along semantic links. After generating the node sequence using the meta-path-guided random walk, a skip-gram model is trained on these sequences to optimize the semantic vector representation of the nodes, emphasizing semantic co-occurrence along the propagation path. It's important to note that the training objective of skip-gram here differs from that of HGNN; the node embeddings generated by HGNN cannot be directly used as a substitute. Training the "node sequence" sampled by skip-gram is a form of local path optimization. Taking the above meta-path as an example, a random walk was performed based on this path, resulting in the following node sequence:

[0154] ;

[0155] express Commenting on the "fourth lane" post creates a weak collaborative chain. Set the window size to 2 and take the center node. Then one of the input samples of the model is The output sample sequence should be:

[0156] ;

[0157] Through iterative training, the learned vectors bring co-occurring nodes closer together. Finally, a co-occurrence vector representation (Embedding2) for each node is obtained.

[0158] Finally, by integrating HGNN feature embedding, Skip-gram collinear embedding (the node feature vector is the concatenation of the HGNN feature embedding and Skip-gram collinear embedding vectors) and edge weights, a richer multidimensional heterogeneous graph (which includes both social and propagation information) is constructed.

[0159] ;

[0160] S14. Account Community Analysis:

[0161] Based on multidimensional heterogeneous graph Graph community detection algorithms (such as Louvain's algorithm) are used to calculate modularity. The higher the value, the more tightly connected the communities are internally, and the more sparsely connected the communities are externally. The goal of community detection algorithms is often to maximize modularity, thereby finding the "most closely cooperating group of nodes," and thus identifying community structures such as "multiple accounts cooperating in posting" and "forwarding content from the same source."

[0162] ;

[0163] in, The edge weight is the edge weight between account nodes i and j (a continuous real number, not 0-1). This represents the weighted degree of node i (i.e., the sum of the weights of all edges connected to i). This represents the total weight of all edges in the graph; This indicates that if nodes i and j belong to the same community, the value is 1; otherwise, it is 0. It is the expected edge weight (i.e., the expected value) under the random network model.

[0164] for example Figure 2 As shown, based on multi-view heterogeneous graphs This leads to several highly relevant robotics communities:

[0165] Step S2: Construction of a two-part graph network of "Account-Post-Entity-Stance"

[0166] Building upon the identified bot communities, a target semantic subgraph network (a bipartite graph network of "account-post-entity-position") is further constructed to analyze the attack targets, positional biases, and content patterns of social bots in a specific semantic space, supporting intent recognition tasks. This mainly includes extracting the target subgraph structure, constructing the "account-post" behavior graph, constructing the "post-entity" semantic graph, constructing the "post-position" semantic graph, constructing the "account-entity-position" semantic graph, and generating the "account-post-entity-position" bipartite graph network.

[0167] S21. Extract the target subgraph structure

[0168] The multi-view heterogeneous graph generated in step S-1 In the process, select a specific social robot community (such as a highly collaborative subgraph with a tight propagation structure), and within that subgraph, based on the initial heterogeneous graph of S-1. Extract nodes and edges related to user accounts, posts, and comments to construct an initial semantic analysis subgraph:

[0169] ;

[0170] ;

[0171] ;

[0172] ;

[0173] This step ensures that the analysis focuses on suspected social bot groups that are "highly collaborative".

[0174] Taking community A as an example, it includes accounts Fifty user nodes, including "Car Soul" and "Road Traveler," were identified. These accounts interacted frequently and shared highly overlapping content themes, initially suggesting they might be a group of social bots from the same source. A total of 230 posts and comments related to these accounts were extracted, forming the initial input for the semantic analysis subgraph.

[0175] S22. Construct an "Account-Post" behavior graph:

[0176] Extract account nodes and their published post nodes from the target subgraph, and establish "account-post" edges. Each edge is assigned a weight to characterize the consistency of the account's behavior towards the content.

[0177] First, from the extracted subgraph Extract the node set and edges:

[0178] ;

[0179] ;

[0180] ;

[0181] Indicates account Posted a message Each edge can have a weight, meaning the weight calculation formula can be defined as follows:

[0182] ;

[0183] in, This indicates the account's influence. This indicates the similarity between the post and the account's previous content (consistency in language style).

[0184] by Taking an account as an example, the account has published a total of 12 posts. Based on the above algorithm, the edges between the account and the posts are weighted, where:

[0185] The post p1 is highly consistent with the account's historical content, with a style similarity of 0.92;

[0186] ...

[0187] Post p12 is a repost, and its style is not completely consistent with its original post, with a similarity of 0.65.

[0188] Therefore, the system forms a graph and edges with the following behavior:

[0189] ( → p1), weight 0.92

[0190] ...

[0191] ( → p12), weight 0.65

[0192] These behavioral patterns form a style map of the account's content expression, which helps in subsequent identification of whether there are consistent language templates or repetitive language manipulations between accounts. This ultimately yields the "Account-Post" behavioral map. This diagram reveals the frequency and patterns of the account's language operations, forming a behavioral background map for intent analysis.

[0193] S23. Construct a semantic graph of “post-entity”:

[0194] Using large language models (such as ChatGPT and DeepSeek) to identify named entities (Entity objects) in text, i.e., targets of post attacks / attention (covering names of people, organizations, countries / regions, event names, topics, etc.), allows... For the entered post content, This represents the set of entities extracted from the text by the large language model:

[0195] ;

[0196] We uniformly use methods such as large language models, business rule base conversion, or embedding vector similarity calculation to disambiguate all entities and generate the final entity nodes.

[0197] Finally, construct the edges from "post nodes" to "entity nodes". , Weights represent the strength or relevance of mentions; confidence scores from the larger model output are used instead.

[0198] ;

[0199] by Taking post p1 as an example, the discussion topic is identified based on the large language model as "Car Company C Auto", with a confidence level of 0.95. Therefore, the following edge is generated:

[0200] (p1 → Car company C), weight 0.95

[0201] If in another post p2, the discussion topic is identified as "Car Company B" with a confidence level of 0.93 based on the large model, then the following edge is generated:

[0202] (p2 → Car Company B), weight 0.93 ......

[0203] By identifying all post entities, the system ultimately builds a complete "post-entity" semantic subgraph, representing the focus of content attention. The final "post-entity" semantic graph is obtained. The corresponding edge is .

[0204] S24. Construct a semantic graph of "post-stance":

[0205] First, we use Large Language Modeling (LLM) to analyze the sentiment stance towards the target entity in the text. (Support, Oppose, Neutral), Order For the entered post content, For probability, This indicates the large language model's judgment of the post's stance:

[0206] ;

[0207] The edge and weight formulas are constructed based on probability, where l can be either supportive, opposed, or neutral:

[0208] ;

[0209] Continuing with the previous example, post p1 (praising car company C) is judged as supporting the stance with a probability of 0.88; post p1 (criticizing car company B) is judged as opposing the stance with a probability of 0.91. The following edges are generated:

[0210] (p1 → Support), weight 0.88;

[0211] (p2 → Oppose), weight 0.91;

[0212] The graph structure between posts and stances depicts the emotional inclination of each text in terms of stance. This ultimately yields a semantic graph of "posts – stances". The corresponding edge is .

[0213] S25. Construct a semantic graph of "Account – Entity – Stance":

[0214] To more accurately model the stance and expression of an account toward a specific entity, the existing "post-entity" and "post-stance" relationships are further integrated into the account dimension to construct a "account-entity-stance" ternary relationship structure, forming a semantic intent graph that directly reflects "which account, which entity, and what attitude it holds".

[0215] First, extract valid triplet paths. Then, iterate through a given account. All post nodes below If it is associated with a certain entity and a certain stance Both have edges:

[0216] ;

[0217] ;

[0218] This means that the post under a certain account expresses the entity's position, thus forming a triple path:

[0219] ;

[0220] To reflect the intensity of an account's sentiment towards a particular entity, the following weighting function is defined:

[0221] ;

[0222] in, For simultaneous accounts All entities mentioned at the same time And express a stance A collection of posts; The intensity of account and post behavior (such as frequency or style similarity). To assess the confidence level of the entities mentioned in the post, The probability score for expressing a stance in a post.

[0223] If the account "Fourth Lane" mentions "Car Company B" multiple times in its posts and expresses an "opposition" stance, the following semantic edges are generated after aggregation:

[0224] ("Fourth Lane", Automaker B, Opposes) → Weight: 0.89;

[0225] Similarly, we can also find:

[0226] ("Fourth Lane", Automaker C, Support) → Weight: 0.78;

[0227] ("Fourth Lane", Automaker D, Neutral) → Weight: 0.45;

[0228] The final result is a semantic graph of "account – entity – stance". .

[0229] S26. Generation of a two-part diagram network: "Account – Post – Entity – Stance":

[0230] The above behavior diagram ( ) and 3 semantic graphs ( , , The data will be integrated to form a complete two-part diagram structure: "Account – Post – Entity – Stance". .

[0231] for example Figure 3 As shown, by constructing a bipartite network diagram, we can clearly see which posts (green nodes) an account (blue node) has published, the corresponding stance of each post (pink node), and the target entities mentioned (yellow node). Through the account's posts, their corresponding stances, and the mentioned entities, we can obtain the stance relationships between the account and the entities. Taking the account "Fourth Lane" as an example, it published 12 posts (4 opposing posts, 5 supporting posts, and 3 neutral posts). The opposing posts involve car company B; the supporting posts involve car company C; and the neutral posts involve car company A. Therefore, the account "Fourth Lane" can be broadly viewed as having a persona of "supporting car company C, opposing car company B, and being neutral towards car company A."

[0232] Step S3: Social Robot Organization Intent Recognition

[0233] Suspicious bot account groups were identified in S1. S-2 further constructed an "account-entity-position" relationship graph to characterize the positional inclination of each account. However, the S-2 relationship graph still suffers from coarse granularity and limited expressive power. For example, it cannot determine whether an account repeatedly speaks out around a specific entity; it cannot distinguish between "occasional opinion expression" and "continuous topic intervention." Therefore, it is necessary to integrate technologies such as graph neural networks (GNN) and large language models (LLM) to identify and predict the organizational intentions of social bot groups and infer whether they exhibit behaviors such as "organized actions" or "common targets." This includes account goal analysis, sentiment and position consistency analysis, collaborative behavior analysis, social network structure and role analysis, and account pattern recognition.

[0234] 1. Account Target Analysis:

[0235] Although a semantic graph of "account-entity-position" has been constructed in S-2, the result is still coarse-grained and difficult to describe in detail the specific semantic targets and potential manipulative intentions of an account. Therefore, it is necessary to introduce a graph neural network model based on the heterogeneous graph of multiple relationships of "account-post-entity-position" to model the semantic path propagation process, identify which objects each account frequently speaks around semantically, and infer its "target of attention" or "target of attack".

[0236] First, based on the two-part diagram structure of "account – post – entity – stance" Extract the relationship structure diagram of "account-post-entity object". .exist The image shows the Large Language Model (LLM) for obtaining post nodes. The semantic embedding vector, i.e., semantic information:

[0237] ;

[0238] Secondly, a heterogeneous graph attention network (HetGAT) model is trained, with the input of the model being... The output is a tag indicating whether the account has attacked or followed a target. Simultaneously, based on the trained HetGAT, semantic information can be propagated from post nodes to account and entity nodes:

[0239] ;

[0240] ;

[0241] That is, after training, and These can effectively represent the semantic attention of an account and the attention received by an entity, respectively. The above formula is a weighted average aggregation function, a core operation in HetGAT. It is used to aggregate features from the node's neighbors to update the representation of the current node. Since different posts have varying degrees of "representativeness" to the account's intent—some posts are emotional, some are ordinary daily life, and some contain critical attacks—HetGAT assigns different weights to each post. :

[0242] ;

[0243] Ultimately, based on account semantic representation (Representing the account's semantic "focus") and the semantic popularity of entity objects (Representing the degree of semantic focus on an entity during discussion), an indirect semantic path can be constructed between "account and entity," thus revealing the intensity of an account's attention to / attacks towards a particular entity:

[0244] ;

[0245] Identify which accounts frequently engage in semantic discussions surrounding a specific entity / topic (i.e., potential attack targets or topic objectives). For example, if a group of accounts consistently engages in discussions surrounding "Car Company B" in real-world data, a semantic path of "Account → Attack → Car Company B" can be constructed. The attack intensity can then be... If the threshold is set to 0.85, and only edges with a value greater than 0.85 are displayed, then a batch of social media accounts that attacked car company B can be obtained.

[0246] 2. Emotional stance consistency analysis:

[0247] In the example above, although the behavioral path of "a certain group of accounts attacking the understanding of cars" has been identified in the account target analysis, it still cannot specifically reveal whether the emotional intensity and stance of these accounts' attack behavior are consistent.

[0248] First, a large language model is used to perform sentiment analysis on the target post text (in this example, the target post text is all posts about "attacking car company B") to obtain sentiment polarity scores (positive, negative, neutral). The fraction values ​​range from -1 to 1:

[0249] ;

[0250] Secondly, if a certain entity object The entity is mentioned in multiple posts. By summing the sentiment scores of these posts, we can obtain the entity's average sentiment stance.

[0251] ;

[0252] in, For all mentioned entities A collection of posts. Therefore This represents the aggregated sentiment value of an entity.

[0253] Therefore, it is possible to calculate a certain account For a certain entity The emotional tendency of the account The post mentioned the entity Calculate the sentiment average of all posts:

[0254] ;

[0255] in, For account All mentioned entities A collection of posts. Therefore This refers to the account's emotional attitude towards the entity.

[0256] Ultimately, for a group of accounts U (such as those attacking car company B), and for a specific entity... The consistency analysis is as follows:

[0257] ;

[0258] If the maximum value is negative. The figure of 85% indicates that 85% of the accounts in the group attacking car company B hold a negative attitude towards car company B. Therefore, the account group U has a highly consistent attack behavior against car company B, indicating that the entity is being attacked in an organized manner by group U.

[0259] In addition, regarding accounts For one or more attacked entities (such as "Car Company B"), after arranging the entities in a unified order, the following sentiment consistency feature vector can be constructed:

[0260] ;

[0261] 3. Collaborative Behavior Analysis:

[0262] In the example above, after account group U expresses consistent emotional inclination and stance towards a certain entity (such as car company B), whether the group exhibits planned and coordinated behavior is a key basis for further identifying whether the social bot group is controlled and whether it is carrying out organized attacks.

[0263] First, based on the heterogeneous graph, for each account u in the group U, obtain the post times of all its posts. Divide the time windows into units such as hours, days, and weeks, and construct them into a time density series:

[0264] ;

[0265] in, Let be the number of posts in the k-th time window.

[0266] time density series Input the data into a time series model (such as LSTM, GRU, Transformer Encoder, etc.) to obtain the behavioral feature representation of the account. This can indicate the account's posting rhythm, periodicity, and whether it posts day and night, among other behavioral characteristics.

[0267] ;

[0268] The time series model needs to be trained. The input to the model during training is a large sequence of time density vectors from various accounts, and the output is a label indicating whether an account belongs to a bot organization. Based on the trained model, the time series data of any account can be analyzed. Mapping to its behavioral representation vector .

[0269] Secondly, for any two accounts in group U Calculate the cosine similarity between their behavioral feature vectors to measure their synchronicity in posting time:

[0270] ;

[0271] Construct an account-account behavior similarity graph based on the cosine of the included angle. , among which the side The weight is the cosine similarity of the included angle, as mentioned above. A similarity threshold is set. Only retain highly synchronized account behavior:

[0272] ;

[0273] Based on high similarity subgraphs Graph clustering algorithms (such as Louvain) were used to identify groups of accounts that were acting in concert.

[0274] ;

[0275] The behavior patterns of accounts within each group are highly similar.

[0276] Finally, determine whether each group is involved in an organized attack. If a certain account group... The following conditions must be met:

[0277] 1) Number of members (Default configuration is 20 people or more);

[0278] 2) Highly consistent attack targets (e.g., all attacking car company B, thus affecting the consistency of emotional stance test) (greater than 0.9)

[0279] 3) The posting rhythm is highly similar (e.g., the behavioral similarity graph structure is dense, that is, the weight of the edges is greater than 0.9).

[0280] This allows us to determine the account group. There is a high probability that these accounts are part of an organized attack on car company B. They likely belong to the same group of coordinated "bot teams" that are systematically and synchronously attacking their target.

[0281] 4. Social Network Structure Role Analysis:

[0282] Based on the account groups discovered above It identifies whether there is an organizational structure, that is, whether there is a central node (commander / initiator), and thus infers who is the commander, who is the responder, who is the core of the action, and who is the peripheral node.

[0283] First, the account collaboration subgraph Identifying groups using Graph Attention Network (GAT) The influence of each node. The graph attention network (GAT) here needs to be pre-modeled and trained; the model input is... The output is also a judgment on whether the accounts belong to the same organization (pre-labeled). The final graph attention network (GAT) is obtained through training, and based on this, attention is applied to each pair of nodes. Attention is calculated for the edges between them, and the core formula is:

[0284] ;

[0285] in, Indicates account The initial embedding can be a concatenation of the time feature vector, content embedding vector, and sentiment consistency vector of the account node; TM (Transformation Matrix) is a trainable linear transformation matrix used to reduce dimensionality or transform the embedding space, and is obtained by training GAT. The attention weight vector is a trainable attention weight vector, which is multiplied by the concatenated vector to obtain the attention score, which is obtained by training the GAT model; This indicates that two vectors have been concatenated; LeakyReLU is the activation function. For nodes Normalize all neighbor nodes j so that ; This is to arrive at the "importance of edges," or "cooperation strength," which is the final learned value. Therefore, the node... The total attention weight is:

[0286] ;

[0287] Similarly, the nodes can be calculated sequentially. Degree centrality (Edges of a node divided by the total number of nodes), Nodes First attack timestamp .

[0288] Therefore, node roles can be assigned based on node degree centrality, attention value, and behavior initiation time:

[0289] 1) Leader

[0290] Commander nodes are generally characterized by: broad connectivity, strong attention, and early posting, i.e., they meet the following conditions:

[0291] ;

[0292] in, The degree centrality threshold (e.g., 0.8). This represents the maximum degree centrality of all accounts in the current account group. If the account... The connectivity is higher than the maximum connectivity in the group. The multiple indicates that it is a "high-central node" in the network; This is the attention weight threshold (e.g., 0.8). The sum of the maximum attention values ​​of all accounts in the group. If an account has a high attention influence value on others in the GAT (exceeding a certain percentage of the group's maximum value), it indicates that it may play an important role in the organization. Allowing for advance planning, The earliest time an attack post was made in the group, if the account The posting time is close to the earliest time in the group, indicating that it may be the initiator or the first wave of attackers.

[0293] 2) Follower

[0294] Responder nodes are generally characterized by: high responsiveness to others, similar content, and accounts that do not proactively initiate collaboration. After determining the leader, a specific account can be calculated. The content similarity with the leader's account in the group was obtained. Then, the following conditions are met:

[0295] ;

[0296] in, For account The number of incoming edges, This is the threshold coefficient (usually set to 0.8), if the account... of Higher than the maximum in the group The multiple indicates that this account is a passive response center, meaning that the account is influenced by many people; For account The number of outgoing edges indicates the degree to which it actively influences others. This is the threshold coefficient (usually set to 0.8), if the account... of Less than the maximum value in the group The fact that the number of times the account was counted indicates that this account did not significantly influence other accounts and was not the leader or the core of the dissemination. The similarity threshold coefficient (usually set to 0.8) is used to determine whether a respondent's content closely mimics the leader's statements, such as using standardized rhetoric or repeating attack slogans. Only when the content is very similar (greater than the threshold) is it considered to be imitating the leader.

[0297] 3) Outer Accounts

[0298] Peripheral account nodes are typically characterized by being isolated, unresponsive, and potentially exhibiting "unorganized" bandwagoning behavior. (This involves) calculating account groups. Average first post time An account can be marked as an external account if it meets any of the following conditions:

[0299] ;

[0300] in, A lower attention weight threshold (e.g., 0.2). This indicates the threshold corresponding to a very low content similarity (e.g., 0.6). This indicates the time delay threshold (e.g., 24 hours).

[0301] 5. Account pattern recognition:

[0302] After identifying the account group Following the roles of "commander" and "responder" in social networks, the next step is to analyze and identify organized behavioral patterns in these accounts. This step integrates the multimodal features (content, sentiment, temporal behavior, and social type) extracted in the previous four steps, aiming to reveal whether there is an intention to organize actions behind them and to make classification predictions.

[0303] First, define the multimodal feature representation vector of account u. It is composed of the following four types of sub-vectors concatenated:

[0304] ;

[0305] in, The content semantic vectors are encoded using a large language model to encode the combination of account posts. This is a vector representing the consistency of emotional stance, derived from emotional stance consistency analysis. The time behavior pattern vector is encoded using a posting time series model; It is a social structure embedding vector, encoded using a graph attention network (GAT) model.

[0306] The fused features Input a heterogeneous graph neural network (HGNN) and learn semantic interactions and collaborative behavior graphs between accounts through an inter-layer attention mechanism. Output is the behavioral pattern label of the account. :

[0307] ;

[0308] Here, the heterogeneous graphical neural network (HGNN) obtains the pattern labels through modeling. Definitions include, but are not limited to:

[0309] Tag 1: Concentrated Attack – The intention is to undermine trust by targeting the same person, organization, or institution with multiple accounts in a concentrated, high-frequency, and consistent manner, thereby weakening its credibility and social standing.

[0310] Tag 2: Public opinion guidance - guiding group cognition, actively creating or spreading specific stances and rhetoric, promoting cognitive deviation, and guiding the group's judgment and emotional direction on specific issues.

[0311] Tag 3: Identity Disguise – Concealing true intentions, confusing account identity and stance (such as impersonating the public or engaging in satirical mockery), interfering with the judgment of information sources, and misleading the public's trust in the source.

[0312] Tag 4: Emotional Manipulation – Disrupting collective cognition by using extreme and anxious emotional content to create panic, anger, and helplessness, thereby undermining public confidence and inducing collective cognitive confusion.

[0313] Based on the above pattern analysis, the pattern tags of a certain account can be accurately identified.

[0314] This system captures heterogeneous data (text, retweet relationships, user profiles, etc.) in real time from large-scale social media platforms (such as Facebook, Twitter, Reddit, etc.), and uses heterogeneous graphs to unify the modeling of various nodes and relationships, including users, topics, and interactive behaviors. This significantly improves the ability to detect hidden manipulation intentions in cross-platform and multilingual public opinion events. Validated on a test set of real online public opinion data, the accuracy rate is 92%-95%, representing an average improvement of approximately 5-10 percentage points compared to traditional detection methods based on single relationship graphs, with particularly significant results in complex scenarios involving cross-topics and cross-platforms. The system can also automatically generate analysis reports on potential manipulation targets and behavioral patterns of social bots, providing valuable references for public opinion analysis for governments, enterprises, and research institutions.

[0315] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A social robot intent recognition system based on heterogeneous graph analysis, characterized in that: It includes a heterogeneous social behavior graph construction and analysis module, an account-post-entity object-position bipartite graph network construction module, and a social robot organizational intent recognition module, which are connected in sequence. The heterogeneous social behavior graph construction and analysis module captures the complex behavioral relationships and network structure characteristics of social robots on the platform by constructing heterogeneous social behavior graphs, embedding social features of account nodes and calculating dynamic edge weights, embedding propagation link features of account nodes, and analyzing account communities, thereby revealing collaboration patterns and obtaining highly collaborative account subgroups. The account-post-entity-stance bipartite graph network construction module extracts the target semantic subgraph structure of the highly collaborative account subgroup, constructs the account-post behavior graph, post-entity semantic graph, post-stance semantic graph, and account-entity-stance semantic graph, and generates the account-post-entity-stance bipartite graph network. By constructing semantic subgraphs, it can achieve a fine-grained understanding of the manipulated objects and stance intentions, and characterize the stance tendencies of each account. The social robot organizational intent recognition module uses a large language model and a heterogeneous graph attention network model to perform account goal analysis, sentiment stance consistency analysis, collaborative behavior analysis, and social network structure role analysis on accounts in the account-post-entity object-position bipartite graph network to obtain multimodal features. The module then performs account pattern recognition, visualization, and path tracing on the multimodal features to identify and predict the organizational behavior patterns and intents of the accounts, and outputs the account behavior pattern labels. The social robot's organizational intent recognition module includes the large language model, the heterogeneous graph attention network model, the time series model, the graph attention network model, and the heterogeneous graph neural network model; The social robot organization intention recognition module extracts an account-post-entity object-relationship structure diagram from the account-post-entity object-stance bipartite network In The semantic embedding vector of the large language model to the post node is obtained in the figure to obtain semantic information; the semantic information is input into the heterogeneous graph attention network model, and the label of whether the account attacks or pays attention to the target is output, and the semantic information is propagated from the post node to the account and the entity node by the heterogeneous graph attention network model. The semantic embedding vector of the large language model to the post node is obtained in the figure to obtain semantic information; the semantic information is input into the heterogeneous graph attention network model, and the label of whether the account attacks or pays attention to the target is output, and the semantic information is propagated from the post node to the account and the entity node by the heterogeneous graph attention network model. The time series model identifies behavioral features based on the time density vector sequence of accounts and outputs behavioral feature vectors; it identifies highly synchronized subgroups based on the similarity of the behavioral feature vectors, thereby obtaining a group of accounts with highly coordinated actions. The graph attention network model identifies the influence of each node in the highly collaborative action account group subgraph; The heterogeneous graph neural network learns the semantic interaction and collaborative behavior graph between accounts in the multimodal features through the inter-layer attention mechanism, and outputs the behavioral pattern labels of the accounts. The multimodal features are feature vectors that integrate content semantic vectors, sentiment consistency vectors, temporal behavior pattern vectors, and social structure embedding vectors. 2.The social robot intent recognition system based on heterogeneous graph analysis of claim 1, wherein: The identification method includes the following steps: S1. Construction and Analysis of Heterogeneous Social Behavior Graph: The heterogeneous social behavior graph construction and analysis module defines node types and relationship types, maps structured data to nodes and edges, and then calculates the social feature embedding vector and dynamic edge weight of the account through social feature embedding and dynamic edge weight. Then, it obtains the propagation link feature embedding vector of the account through the propagation link feature embedding of the account node. Based on the social feature embedding vector, the dynamic edge weight and the propagation link feature embedding vector, community detection is performed to obtain the highly collaborative account subgroup. S2, the account-post-entity object-stance bipartite graph network construction module extracts relevant nodes and relationships from the high coordination account sub-group, performs structure extraction of the target semantic sub-graph, edge structure establishment to obtain the account-post behavior graph, identifies post named entities and confidence, establishes edge weight to obtain the post-entity semantic graph, analyzes post stance score to establish stance edge to obtain the post-stance semantic graph; construct the ternary relationship of account, entity and stance and calculate the weight to obtain the account-entity-stance semantic graph, and then fuse the account-post behavior graph, the post-entity semantic graph, the post-stance semantic graph and the account-entity-stance semantic graph to obtain the account-post-entity object-stance bipartite graph network. The account-post-entity object-stance bipartite graph network is used to analyze the attack object, stance bias and content mode of social robots in specific semantic space, and supports the intent recognition task; S3, the social robot organization intent recognition module performs account target recognition using a large language model and a heterogeneous graph attention network, identifies a group of accounts that frequently speak around a target entity; then, based on the identified group of accounts, perform sentiment stance consistency analysis, analyze the sentiment polarity of the group accounts, aggregate the entity emotion scores of the group accounts, and determine whether the group has consistency in stance and emotion towards the target entity; then, for the group with consistent stance and emotion towards the target entity, perform coordinated behavior analysis, calculate the behavior synchronization degree between accounts, identify a highly synchronized subgroup as the basis for judging the group of organized attacks, and output a high coordination action account group; then, analyze the social network structure of the high coordination action account group, and identify the role of the account; perform account pattern recognition on the identified role accounts, determine the intent of the account pattern; finally, map the intent recognition result to the graph for visual display and path tracing, and output a graphical view and report friendly to supervisors and behavior pattern labels of accounts. 3.The social robot intent recognition system based on heterogeneous graph analysis of claim 2, wherein: Step S1 includes the following steps: S11, heterogeneous social behavior graph construction: the heterogeneous social behavior graph construction and analysis module defines the node types and relationship types of the heterogeneous social behavior graph based on the collected account information and post information after using a large language model to encode the post text, and obtains a heterogeneous graph G, which includes nodes V and edges R; S12, account node social feature embedding and edge weight calculation: the heterogeneous social behavior graph construction and analysis module uses a heterogeneous graph neural network for feature embedding, maps the information of each account node in the heterogeneous graph G into a low-dimensional social feature embedding vector, calculates a dynamic edge weight according to content similarity, time proximity and account interaction frequency using a dynamic edge weight calculation mechanism, the dynamic edge weight is the cooperative relationship weight between accounts, and the account social heterogeneous graph is obtained according to the social feature embedding vector and the dynamic edge weight , the account social heterogeneous graph includes account node embedding features and cooperative relationships; S13, account propagation link feature embedding: the heterogeneous social behavior graph construction and analysis module first generates a node sequence based on meta-path guided random walk, then trains, learns and optimizes the semantic vector representation of the nodes in the node sequence by using a skip word model, emphasizes the semantic co-occurrence relationship in the propagation path, obtains a Skip-gram co-linear embedding vector as a propagation link feature representation of the account, and then fuses the Skip-gram co-linear embedding vector with the account social heterogeneous graph to obtain a multi-dimensional heterogeneous graph that simultaneously includes social structure information and propagation information ; S14, using a graph community discovery algorithm to find the sub-group of nodes with the highest modularity score based on the multi-dimensional heterogeneous graph computing and maximizing modularity finding the most tightly-knit group of nodes, identifying the community structure of multi-account coordinated publishing and homogenous content forwarding, and obtaining the sub-group of high coordination accounts 4.The social robot intent recognition system based on heterogeneous graph analysis of claim 3, wherein: In step S11, the node types include accounts, posts and comments, and the relationship types include posting / liking relationships, quoting / forwarding relationships and commenting / liking relationships. The edges in the relationship types all include timestamp information; In step 12, the social feature embedding vector represents the identity representation of the current node after aggregating all semantic and structural information in the heterogeneous graph G; the content similarity is the cosine similarity, the time proximity is an exponential decay function, and the interaction frequency between accounts includes forwarding, commenting and liking; in step S13, the skip-gram model learns vectors through continuous iterative training, which makes the distance between co-occurring nodes closer, and obtains a co-occurrence vector representation of each node. The Skip-gram co-linear embedding vector is a spliced vector of the social feature embedding vector and the co-occurrence vector representation; In step S14, the graph community discovery algorithm is Louvain algorithm, and the modularity Q is a function of edge weight between two accounts, node weighted degree, total weight of all edges, and edge weight expected in a random network model. The greater the value, the closer the community is internally connected, and the sparser the community is. 5.The social robot intent recognition system based on heterogeneous graph analysis of claim 2, wherein: Step S2 includes the following steps: S21, the account-post-entity object-stance two-part graph network construction module selects a specific social robot community from the high coordination account sub-group and extracts relevant nodes and edges within the sub-graph range to obtain an initial semantic analysis sub-graph ; S22, extracting a node set and an edge from the initial semantic analysis subgraph each edge being attached with a weight, to obtain the account-post behavior graph, the account-post behavior graph revealing the language operation frequency and mode of the account, and constituting a behavior bottom graph for the intention analysis; S23, using a large language model to identify named entities in the post and extract entity sets, disambiguate all entities, generate final entity nodes, use the confidence level output by the large language model to construct edges from post nodes to entity nodes, and obtain the post-entity semantic graph; S24, using a large language model to analyze the sentiment of the target entity in the post obtain the post-entity semantic graph; S24, using a large language model to analyze the sentiment of the target entity in the post obtain the post-entity semantic graph; S24, using a large language model to analyze the sentiment of the target entity in the post S25, integrate the Weibo-entity semantic graph and the Weibo-stance semantic graph into the account dimension, extract effective triple paths to construct the account-entity-stance semantic graph, and form an intent graph in semantics; S26, fuse the account-Weibo behavior graph, the Weibo-entity semantic graph, the Weibo-stance semantic graph and the account-entity-stance semantic graph to obtain the account-Weibo-entity object-stance bipartite graph network, and determine the Weibo published by the account, the corresponding stance condition and the target entity mentioned by the Weibo, and obtain the stance relationship between the account and the entity through the account Weibo corresponding stance and the mentioned entity. 6.The social robot intent recognition system based on heterogeneous graph analysis of claim 5, wherein: In step S21, the specific social robot community is a high-collaboration subgraph with a tight propagation structure; In step S22, the weight of each edge is a function of the account influence and the similarity of the Weibo and the historical content of the account; In step S23, the large language model is ChatGPT or DeepSeek, and the named entity is the target of Weibo attack / attention, including name, organization name, country / region, event name and topic; disambiguation processing is performed using a large language model, a business rule base conversion or an embedding vector similarity calculation method; and a confidence degree is used to replace the weight representing the mention intensity or correlation; In step S24, the sentiment stance including support, opposition and neutrality, the post-stance semantic graph depicts the tendency of each piece of text in the stance sentiment; In step S25, all Weibo nodes under the account are traversed, and if a Weibo node has edges with an entity and a stance, it is determined that the Weibo under the account expresses the stance of the entity, and a triple path is formed: The weight function reflecting the intensity of the sentiment expressed by the account towards the entity is the collection of posts that mention the entity and the account expressing a stance The function of the behavior intensity of the Weibo, the confidence degree of the Weibo mentioning the entity and the probability score of the Weibo expressing the stance, and the behavior intensity of the Weibo includes frequency and style similarity.

7. The social robot intent recognition system based on heterogeneous graph analysis of claim 2, wherein: In step S3, when identifying the account target, the heterogeneous graph attention network model aggregates features from node neighbors by weighted average aggregation, updates the representation of the current node, and assigns different weights to each Weibo based on the account semantic representation, constructs an indirect semantic path between the account and the entity based on the account semantic representation and the semantic heat of the entity object, and obtains a content semantic vector, the account semantic representation is the focus of the account in semantics, and the semantic heat of the entity object is the semantic focus degree of the entity being discussed; When analyzing the consistency of emotional stance, a large language model is called to perform sentiment analysis on the target Weibo text to obtain a sentiment polarity score, the sentiment polarity score is between -1 and 1, including positive, negative and neutral; When an entity is mentioned by multiple posts, the aggregated post sentiment scores are averaged to obtain the entity's average sentiment stance; The average emotion of all Weibo mentioning the entity in the Weibo of the account is obtained to obtain the emotional tendency of the account to the entity: Finally, a consistency analysis of the entities of a group of accounts U is performed: consistency analysis of the entities of a group of accounts U is performed: 。 8.The social robot intent recognition system based on heterogeneous graph analysis of claim 2, wherein: In step S3, in the collaborative behavior analysis, based on multi-dimensional heterogeneous graph For each account u in the group U, get all the post timestamps and divide the time window by time unit to get the time density vector sequence: ； wherein, is the number of posts for the kth time window; A sequence of time density vectors Input the behavior feature vector of the account to the time series model to obtain a highly synchronized sub-group, the behavior features including the writing rhythm, periodicity, and whether the account is day and night. According to the number of members in the high-synchronization subgroup, the probability of emotional stance consistency test and the size of edge weight value, it is determined whether each group exists organized attack, and a high-collaboration account group is output. 9.The social robot intent recognition system based on heterogeneous graph analysis of claim 2, wherein: In the social network structure role analysis, the influence of each node in a high coordination action account subgraph is identified by using a graph attention network model to obtain coordination strength ; ； in, For account The initial embedding is a concatenation of the temporal feature vector, content embedding vector, and sentiment consistency vector of the account node; TM is a trainable linear transformation matrix used for dimensionality reduction or transformation of the embedding space, which is obtained by training the graph attention network model. Let be a trainable attention weight vector, and The attention score is obtained by multiplying the results, which is trained from the graph attention network model. This indicates that two vectors have been concatenated; LeakyReLU is the activation function. For nodes Normalize all neighbor nodes j, such that ; Node The total attention weight for the node ; Divide the number of edges of a node by the total number of nodes to get the number of nodes. Degree centrality and obtain the node First attack timestamp ; Based on degree centrality , attention value , first attack behavior timestamp Divide node roles: When the following conditions are met, the node role of the account is determined to be a director node: ； wherein, is a threshold for the degree centrality, is the maximum degree centrality of all accounts in the current account group, is a threshold for the attention weight, ≥ 0.7, is the maximum sum of attention values of all accounts in the group, is a time advance, is the time of the earliest attack post initiated in the group; When the following conditions are met, the node role of the account is determined to be a responder: ； wherein, is the in-degree of the account, is the in-degree threshold coefficient, is the out-degree of the account, is the out-degree threshold coefficient, represents the extent to which it actively influences others, is the out-degree threshold coefficient, is the number of accounts in the group, is the number of accounts in the group that are commanders Content similarity of node accounts, is a similarity threshold coefficient; When the following conditions are met, the node role of the account is determined to be a peripheral account: ； wherein, is a lower attention weight threshold, ≤ 1 / 3 , is a threshold value corresponding to a content similarity being too low, is an account the average first post time of the group, is a time delay threshold; The behavior mode label of the account includes: concentrated attack, public opinion diversion, identity camouflage and emotional manipulation.

Images