Comment word frequency statistical method and device based on Internet search and computer storage medium

Abstract

The invention discloses a comment word frequency statistical method and device based on Internet search and a computer storage medium. The method comprises the steps of obtaining comment data including a comment text, a reviewer identity identifier, comment release time and a comment object identifier; performing cleaning, word segmentation and intention recognition on the comment text, generating an intention containing a standardized intention category and a core keyword, and generating node vector representation fusing semantic and structural information; clustering analysis is carried out based on the embedded vector of the reviewer node, and abnormal clusters conforming to the water army behavior characteristics are identified; an identification result is fed back to a data preprocessing link, and the system identification capability is optimized; the method overcomes the rhus and superficial features of a traditional method, further provides abundant and structured input for the learning of a subsequent graph neural network, and enhances the accuracy and reliability of water army recognition.

Description

Technical Field

[0001] This invention relates to comment word frequency statistics, and more particularly to a method, apparatus, and computer storage medium for comment word frequency statistics based on Internet search. Background Technology

[0002] In the field of internet content security governance, identifying online trolls has always been a challenging task. Traditional methods for detecting online trolls usually rely on setting thresholds for the behavioral characteristics of individual accounts. These methods are often ineffective when facing highly organized and coordinated online troll groups, as their members can circumvent single-dimensional detection mechanisms through decentralization of behavior and linguistic camouflage. Especially when online troll groups adopt real user behavior patterns and use natural language generation technology to forge content, identification models based on independent feature analysis are prone to misjudgment or missed detection, seriously affecting the practicality and reliability of governance effectiveness.

[0003] As the operational strategies of online trolls evolve, their collaboration and stealth capabilities are increasing, necessitating the development of a new identification technology capable of analyzing group collaboration. In recent years, comment frequency statistics based on internet search have gained attention. This method crawls and statistically analyzes comment content related to a target topic across multiple platforms within a specific time period, constructing a large-scale corpus and extracting high-frequency words and their frequency distribution features. Compared to traditional methods, this technology does not rely on single-user behavior analysis but starts from macro-level language usage patterns, identifying abnormally concentrated word combinations and their statistical distribution deviations, thereby detecting potentially collaborative online troll groups. This type of method combines the advantages of natural language processing and big data analysis, possessing stronger generalization and anti-disguise capabilities, providing a new approach to improving the accuracy and robustness of online troll identification. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a method, apparatus, and computer storage medium for statistical analysis of comment word frequency based on Internet search.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a comment word frequency statistics method based on Internet search, comprising the following steps:

[0006] S1: Obtain comment data from internet platforms, including comment text, comment posting time, publisher ID, factual identifiers of the comments, and information about the platform from which the comments originated;

[0007] S2: Clean, segment, and identify the intent of the comment text to generate intents containing standardized intent categories and core keywords; anonymize the commenter's identity and extract behavioral features.

[0008] S3: Construct a heterogeneous graph containing comment nodes, commenter nodes, and intent nodes and their corresponding relationships;

[0009] S4: Use heterogeneous graph neural networks to perform embedding learning on heterogeneous graphs to generate node vector representations that integrate semantic and structural information;

[0010] S5: Cluster analysis based on the embedding vectors of commenter nodes to identify abnormal clusters that conform to the characteristics of online troll behavior;

[0011] S6: Feedback the recognition results to the data preprocessing stage to optimize the system's recognition capabilities.

[0012] In a preferred embodiment of the present invention, in S2, the specific operation of generating an intent containing standardized intent categories and core keywords is as follows:

[0013] S21: Perform noise reduction, word segmentation, and stop word removal on the comment text;

[0014] S22: Hash the commenter's identity and extract their historical behavioral characteristics;

[0015] S23: Map comment text to standardized intent categories using a pre-trained intent classification model;

[0016] S24: Extract keywords related to the intent to form a standardized intent.

[0017] In a preferred embodiment of the present invention, the intent classification model is specifically based on the Transformer architecture, and outputs the confidence level of the text belonging to each preset intent category.

[0018] In a preferred embodiment of the present invention, in S3, the correspondence relationship is specifically as follows: a commenter node-comment node edge is created based on the comment posting relationship; a comment node-intent node edge is created based on the comment-intent mapping relationship; and a commenter node-comment node edge is created based on the common intent or behavioral pattern similarity among commenters.

[0019] Commenter nodes and comment node edges represent the comment posting relationship;

[0020] The edge weights between comment nodes and intent nodes are determined by the confidence level of intent recognition.

[0021] Commenter nodes and commenter node edges are created and weighted based on the frequency of shared intentions or the similarity of behavioral characteristics among commenters.

[0022] In a preferred embodiment of the present invention, in S4, the heterogeneous graph neural network includes node-level and semantic-level attention mechanisms, supporting information aggregation of multiple types of nodes and edges.

[0023] In a preferred embodiment of the present invention, in S5, a density clustering algorithm is used to identify abnormally dense clusters among commenter nodes, and this is combined with the meaning within the clusters. Figure 1 Identifying online trolls based on similarity in behavior and other indicators.

[0024] Internet search-based comment word frequency statistics device, including:

[0025] The data acquisition unit is used to acquire comment data from internet platforms. The comment data includes comment text, comment posting time, commenter identification, the identifier targeted by the comment, and information about the platform from which the comment originated.

[0026] The comment data preprocessing and semantic normalization unit is configured to perform preprocessing and semantic normalization operations on the acquired comment data to generate standardized intents;

[0027] Heterogeneous graph construction unit, used to construct a heterogeneous graph based on preprocessed and semantically normalized data;

[0028] The graph neural network processing unit is used to input the constructed heterogeneous graph into the graph neural network model for deep representation learning in order to generate the embedding vectors of the nodes.

[0029] The water brigade identification unit identifies suspected water brigade groups based on node embedding vectors generated by a graph neural network model; and

[0030] The feedback optimization unit is used to feed back the identified suspected troll groups to the comment data preprocessing and semantic normalization stage to continuously optimize system performance.

[0031] In a preferred embodiment of the present invention, the online troll identification unit includes:

[0032] The feature extraction module is used to extract cluster recognition features from the commenter node embedding vector;

[0033] The clustering module is used to identify abnormally dense clusters using a density clustering algorithm;

[0034] The decision module is used to determine the meaning within the cluster. Figure 1 Identifying online trolls based on consistency and behavioral similarity indicators.

[0035] In a preferred embodiment of the present invention, the graph neural network unit includes a heterogeneous graph neural network with node-level and semantic-level attention mechanisms, supporting information aggregation of multiple types of nodes and edges.

[0036] In a preferred embodiment of the present invention, the intent recognition module is configured as a deep learning model based on the Transformer architecture, and outputs the confidence level of the text belonging to each preset intent category.

[0037] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0038] (1) This invention is a method, device and computer storage medium for statistical analysis of comment word frequency based on Internet search. By constructing a heterogeneous graph containing comment nodes, commenter nodes and intent nodes, and introducing commenter-commenter association edges based on common intent frequency or behavioral similarity, the system can transform the collaborative behavior of water army gangs within a specific time window, such as the concentrated release of comments with the same intent, into high-weight connection edges in the graph structure, thereby visualizing the implicit collaborative pattern and realizing the quantitative representation of the collaborative relationship of water army. This makes the association between gang members computable and analyzable. Compared with the existing technology that relies on independent behavior thresholds, the latter cannot capture the many-to-many relationship between commenters. This invention directly models collaboration through graph structure, overcomes the superficiality of traditional methods, and further provides rich and structured input for subsequent graph neural network learning, enhancing the accuracy and reliability of water army identification and more effectively revealing organized water army activities.

[0039] (2) This invention provides a method, apparatus, and computer storage medium for statistical analysis of comment word frequency based on Internet search. It adaptively aggregates heterogeneous information through node-level and semantic-level attention mechanisms to generate node vector representations that integrate semantic and structural information. This enables GNNs to deeply explore collaborative patterns in the graph, such as the dense connections of commenter nodes in the embedding space. The directly generated low-dimensional vectors can characterize the complex attributes and topological positions of nodes, thereby effectively identifying the meaning. Figure 1 This invention identifies anomalous clusters with similar behaviors. In contrast to existing technologies based on simple clustering or rules, which cannot handle relationships between multiple types of nodes and edges, this invention achieves high-precision detection of collaborative behavior through deep learning. This further enhances the system's ability to detect hidden bots, reduces false positives and false negatives, and ensures the comprehensiveness and robustness of the detection.

[0040] (3) This invention is a method, device and computer storage medium for statistical analysis of comment frequency based on Internet search. By feeding back the results of suspected troll groups to the data preprocessing stage, the feature extraction rules are dynamically optimized, so that the system forms a self-learning closed loop. It can adjust the detection focus in real time according to the changes in the troll group's collaborative strategy, which improves the system's adaptability and anti-avoidance ability. It can continuously deal with the new collaborative methods of troll groups. Compared with the static model with fixed rules in the prior art, the latter is easy to become outdated and requires manual intervention. However, this invention achieves dynamic evolution through the feedback mechanism. A further effect is that it forms a long-term effective defense system, reduces maintenance costs, and improves the system's sustainability and practical value. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of a preferred embodiment of the present invention;

[0043] Figure 2 This is the standardized intent category and core keyword intent generation of the preferred embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0046] Application Overview:

[0047] Traditional methods for identifying online trolls mainly rely on threshold judgments of individual comments or user behavior characteristics, such as keyword frequency and posting intervals. This approach is limited to analyzing each comment entity or commenter as an isolated individual. However, with the evolution of online troll tactics, their methods have become increasingly professional and collaborative. By distributing posts, rotating accounts, avoiding sensitive words, and fabricating behavioral patterns, they make individual behaviors appear normal when evaluated independently, thus bypassing traditional detection mechanisms.

[0048] Traditional identification methods lack effective mechanisms to discover hidden connections and collaborative patterns between different commenters and comment content. They are unable to see the common malicious intent and gang manipulation characteristics hidden behind these seemingly independent normal behaviors from a structural level, resulting in a significant decrease in identification performance when facing highly disguised and highly coordinated online trolls.

[0049] This invention provides a method, apparatus, and computer storage medium for statistical analysis of comment frequency based on Internet search, aiming to solve the problems of low accuracy, high false alarm rate, and high false negative rate in the existing technology for identifying highly disguised and highly coordinated online troll groups.

[0050] After introducing the basic principles of the present invention, various non-limiting embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] Exemplary method:

[0052] Figure 1 The illustration shows a flowchart of a comment word frequency statistics method based on Internet search according to an embodiment of this application.

[0053] like Figure 1 The comment word frequency statistics method based on Internet search according to the embodiments of this application includes:

[0054] S1: Obtain comment data from internet platforms, including comment text, comment posting time, publisher ID, factual identifiers of the comments, and information about the platform from which the comments originated;

[0055] S2: Clean, segment, and identify the intent of the comment text to generate intents containing standardized intent categories and core keywords; anonymize the commenter's identity and extract behavioral features.

[0056] S3: Construct a heterogeneous graph containing comment nodes, commenter nodes, and intent nodes, and associate commenters, comments, and intents;

[0057] S4: Use a neural network formed by heterogeneous graphs to perform embedding learning on heterogeneous graphs, and generate node vector representations that integrate semantic and structural information;

[0058] S5: Cluster analysis based on the embedding vectors of commenter nodes to identify abnormal clusters that conform to the characteristics of online troll behavior;

[0059] S6: Feedback the recognition results to the data preprocessing stage to optimize the system's recognition capabilities.

[0060] This invention identifies highly deceptive online trolls by constructing a heterogeneous graph that integrates comments, commenters, and their standardized intentions. Its core principle lies in transforming seemingly isolated behavioral information into a network graph rich in connections. This graph treats comments, commenters, and intentions as three types of nodes, connected through posting relationships, attribution relationships, and key commenter-related edges. The connections between commenters are established based on their posting of comments with the same intention or exhibiting highly similar behavioral patterns, thus effectively capturing potential collaborative signals.

[0061] By using graph neural networks (GNNs) to perform deep learning on heterogeneous graphs, the model adaptively aggregates multi-hop neighbor information through node-level and semantic-level attention mechanisms, learns the complex structure and semantic relationships in the graph, and generates a low-dimensional embedding vector for each commenter that integrates its own attributes, local neighborhood state, and global graph context information.

[0062] In the embedding vector, behavioral coordination, intention Figure 1 Online commentators who engage in online trolling will naturally cluster together to form dense groups; then, density-based clustering algorithms are used to identify these anomalous clusters, and a comprehensive assessment of their internal behavioral similarities and intentions is conducted. Figure 1 By using multiple characteristics such as consistency and abnormal release patterns, highly disguised coordinated online troll groups can be accurately identified.

[0063] The following will describe each step in detail.

[0064] S1: Obtain comment data from internet platforms, including comment text, comment posting time, publisher ID, factual identifiers of the comments, and information about the platform from which the comments originated;

[0065] The acquisition of comment data is achieved by executing web crawling tasks. These tasks are configured with crawling rule sets for specific target websites. The rule sets define the webpage structure parsing path, data field extraction rules, and crawling frequency and depth strategies to ensure the comprehensiveness and timeliness of data acquisition. The specific target websites are e-commerce platforms, social media, forums, and news comment sections.

[0066] The webpage structure parsing path uses XPath or CSS selector expressions to precisely specify the comment text, comment publication time, publisher ID, comment fact identifier, and comment source platform information; the comment fact identifier is the identifier of the product / service / event that the comment addresses. Data field extraction rules are used for formatting the extracted content, such as uniformly converting timestamps to UTC format and uniformly encoding comment text to UTF-8. The crawling frequency uses randomized request intervals, for example, a Poisson distribution random delay of 2 to 8 seconds. The crawling depth strategy defines the upper limit of the number of comment pages loaded for each product page or post page to balance data volume and crawling efficiency.

[0067] In this invention, the data collected—including comment text, comment posting time, poster ID, factual identifiers of the comments, and source platform information—constitute a multi-dimensional analytical framework. This framework comprehensively characterizes comment activity from multiple levels, including content, behavior, identity, and context, providing a solid data foundation for further in-depth analysis of hidden collaborative relationships among users and identification of highly deceptive online troll groups.

[0068] The review text is the core of the analysis, directly carrying the user's intentions and emotions; the release time is used to reveal the temporal patterns and outbreak modes of the reviews and is the key to identifying coordinated review brushing behaviors; the publisher ID is the basis for associating all the behaviors of the same user and constructing their behavior portraits; the review fact identifier is used to aggregate all the feedback under a specific target and analyze the focus of public opinion; the source platform information is used to understand the characteristics of user groups and content differences on different platforms.

[0069] S2: Clean, segment, and identify the intentions of the review text, generate intentions containing standardized intention categories and core keywords, anonymize the reviewers' identities, and extract behavioral characteristics;

[0070] Perform preprocessing and semantic normalization operations on the obtained review data to generate standardized intentions. The preprocessing and semantic normalization operations of the review data specifically include:

[0071] Step S210: Text cleaning and segmentation;

[0072] Perform unstructured data cleaning on the review text to remove HTML tags, special characters, URL links, and emoticons; among them, the method of removing HTML tags includes using regular expressions and HTML parsing libraries to remove the possible HTML tags in the review text; removing special characters includes but is not limited to Japanese emoticons, network symbols, and non-standard Unicode characters.

[0073] After the unstructured data cleaning in the present invention, a Chinese word segmentation model based on the conditional random field (CRF) is used to perform Chinese word segmentation on the review text, obtaining a series of word tokens, and removing the words in the preset stop word set.

[0074] Among them, the CRF model can accurately identify the word boundaries in the review text by learning the sequential labeling relationships between words; the training corpus of the model covers general domain texts and specific e-commerce and social media review domain corpora to improve the word segmentation accuracy of network new words and specific domain terms.

[0075] After word segmentation, the system will remove the words in the preset stop word set; among them, the stop word set contains commonly used but semantically meaningless words, such as "de", "shi", "le", "yi ge", etc., with a scale of more than 2000, and is dynamically maintained and extended according to the actual application scenarios. Removing stop words helps reduce the text dimension and focus on the core semantic information.

[0076] Step S220: Reviewer information extraction and anonymization;

[0077] The commenter's identity is extracted from the comment data and hashed to anonymize it, generating a unique and irreversible anonymous commenter ID. At the same time, metadata related to the commenter, such as the number of historical comments, the average comment posting interval, the number of users followed, and the number of followers, is extracted as the initial set of commenter features. The identity identifier includes user ID, username, or unique session identifier.

[0078] After extracting the commenter's identity, the system performs hash anonymization on the identity to protect user privacy and ensure the compliance of data processing. Specifically, hash anonymization uses the SHA-256 hash algorithm, conforming to the FIPS 180-4 standard, to perform one-way encryption on the original identity, generating a unique and irreversible anonymous commenter ID.

[0079] During the hashing process, the system also extracts metadata related to the commenter as the initial set of commenter features. This metadata is obtained by aggregating and statistically analyzing the commenter's historical behavior data. This metadata is normalized to the [0,1] interval and serves as a component of the commenter node's initial feature vector. The metadata includes: the total number of the commenter's historical comments (e.g., the number of comments posted in the past 6 months), the average comment posting interval (e.g., the standard deviation of comment posting timestamps in the past 30 days), the number of users followed, the number of followers, and the IP address change entropy value (calculated by analyzing the complexity of the distribution of IP addresses used to post different comments; a high entropy value may indicate frequent IP address changes and spoofing).

[0080] Step S230: Intent recognition and standardization;

[0081] Semantic analysis is performed on the cleaned and segmented comment text to identify the core intent conveyed by the comment. Intent recognition is achieved by deploying a pre-trained deep learning model based on the Transformer architecture, trained on a large amount of labeled corpus. This model maps input text to a predefined set of standardized intent categories, such as purchase intent, recommendation intent, consultation intent, complaint intent, support intent, and opposition intent. The set of standardized intent categories is predefined, and each intent category is associated with a unique intent identifier and a set of keywords describing its semantic core. For comments with multiple intents, the model outputs multiple intent categories and their corresponding confidence scores. This invention selects the intent with the highest confidence score above a preset threshold T1 as the primary intent of the comment and standardizes it into the corresponding intent.

[0082] The identification of core intent is achieved by deploying a pre-trained deep learning model. The deep learning model is built on the Transformer architecture and includes a multi-head self-attention mechanism and a feedforward network layer. The model is trained and fine-tuned on Chinese comment corpus, which covers multiple domains such as e-commerce comments and social media discussions.

[0083] The model's output layer is designed as a multi-class classifier to map the input text to predefined standardized intent categories. These standardized intent categories include purchase intent (e.g., want to buy, place an order), recommendation intent (e.g., worth recommending, share good products), inquiry intent (e.g., how is the quality, are there any discounts), complaint intent (e.g., it's terrible, bad experience), support intent (e.g., agree, support), opposition intent (e.g., disagree, bad review), and advertising intent (e.g., contact information, follow me). Each intent category is associated with a unique intent identifier and a set of keywords that describe its semantic core.

[0084] For comments with multiple intentions, the deep learning model outputs each intention category and adds a confidence score to each intention category. The intention with the highest confidence score that is higher than the preset threshold T1 is selected as the core intention of the comment. If the highest confidence score fails to reach the threshold T1, the comment may be marked as having an unclear intention or be subject to further manual review.

[0085] Step S240: Generate standardized intent;

[0086] Based on the standardized intent identified in step S230, one or more standardized intents are generated for each comment. The standardized intent includes an intent identifier, intent type, intent confidence, and a set of core keywords related to the intent extracted from the comment text. The extraction of the core keyword set is achieved by combining part-of-speech tagging results with a text keyword extraction algorithm to select words with actual semantic meaning, such as verbs, adjectives, and nouns.

[0087] The comment text is tagged with parts of speech. Specifically, the identified nouns, verbs, and adjectives with actual semantic meaning are used as input, and a text keyword extraction algorithm, namely the TextRank algorithm, is applied. The TextRank algorithm constructs a word co-occurrence graph and calculates the PageRank of the words in the graph to evaluate the importance of each word. Finally, words that rank highly and meet the part-of-speech requirements are selected as the core keywords for the intended message.

[0088] S3: Construct a heterogeneous graph containing comment nodes, commenter nodes, and intent nodes, and associate commenters, comments, and intents;

[0089] Construct a heterogeneous graph containing comment nodes, commenter nodes, and standardized intent nodes. Specific construction methods include:

[0090] Step S310: Node creation;

[0091] Create a comment node C, a commenter node U, and a standardized intent node I, respectively.

[0092] Here, comment node C represents a specific comment. Its initial feature vector is formed by concatenating the word embeddings of the comment text with the comment metadata. The word embeddings of the comment text are implemented by vectors generated by the BERT model, including the comment metadata.

[0093] Commenter node U represents an anonymous commenter ID. Its initial feature vector consists of commenter metadata extracted in step S220. The commenter metadata includes the number of historical comments, the average comment posting interval, the number of users followed, the number of followers, and the IP address change entropy value. This metadata is further reduced in dimensionality through principal component analysis (PCA) or directly used as the original feature vector.

[0094] A standardized intent node I represents an abstract standardized intent category. Its initial feature vector consists of the embedding vector of that intent category and may include its current statistical frequency in the comment data as an additional feature.

[0095] In the above, the word embedding of comment node C is obtained through a pre-trained BERT model. Specifically, the cleaned and segmented comment text is input into the last layer of the BERT model, and the vector corresponding to the CLS tag is extracted as the semantic representation of the comment. The comment metadata includes comment length, comment sentiment score, and comment publication timestamp. This metadata is normalized by Min-Max and concatenated with the word embedding vector to form the initial feature vector of comment node C. The comment sentiment score is calculated by an independent sentiment analysis model and ranges from -1 to 1. The comment publication timestamp is encoded into sine and cosine components through Fourier transform, or converted into a relative value from a certain reference time and normalized.

[0096] In the above, the embedding vector of the intent category in the standardized intent node I is obtained by Word2Vec or BERT encoding the intent name (such as purchase intent), or by average pooling the keyword embedding vector in the set of core keywords associated with it. In addition, the intent node feature also includes its current statistical frequency in the comment data as an additional feature. This frequency value is logarithmically transformed and normalized. The statistical frequency is specifically the total number of times the intent has appeared in the past 24 hours.

[0097] Step S320: Edge creation and weight assignment;

[0098] Based on the three nodes created above—comment node, commenter node, and standardized intent node—three types of edges are created: commenter-comment edge (UC edge), comment-intent edge (CI edge), and commenter-commenter association edge (UU edge).

[0099] Specifically, the commenter-comment edge (UC edge) is created for each comment, with a directed edge pointing from the commenter node (U) corresponding to its publisher to the comment node (C) corresponding to the comment itself. This directly reflects the attribution relationship of the comment, indicating a direct and definite publishing relationship. The comment-intent edge (CI edge) is created for each comment. If one or more standardized intents are identified, a directed edge is created from the comment node (C) to the corresponding standardized intent node (I). The created comment-intent edge is set with the confidence score of the standardized intent, which directly reflects the strength of the association between the comment and the specific intent. For example, if the confidence score of a comment being identified as having a purchase intent is 0.85, then the corresponding CI edge weight is 0.85. The commenter-commenter association edge (UU edge) is created based on the potential collaborative patterns or similar intents between commenters.

[0100] The aforementioned methods of creating related edges establish connections based on shared intent and behavioral patterns.

[0101] Specifically, the shared intent association means that when two different commenters publish at least K comments within a preset time window T2, and at least L of these comments are identified as having the same standardized intent, then an undirected association edge is created between the two commenter nodes.

[0102] Specifically, behavioral pattern association involves creating an undirected association edge between two different commenters if, within a preset time window T3, the Euclidean distance of their behavioral pattern features (such as the standard deviation of their comment posting time interval, average comment length, and frequency of specific sentiment words in their comments) is less than a preset threshold D1. The behavioral pattern features include: statistical characteristics of the comment posting time interval sequence, average comment character length, frequency of specific sentiment words in the comments, and frequency of mentions of product / service names in the comments. Statistical features include mean, standard deviation, skewness, and kurtosis. Specific sentiment words include positive and negative terms. The feature vectors are standardized, and the Euclidean distance is calculated. The confidence score of the edge is set as the reciprocal of the behavioral pattern similarity, i.e., 1 / (Euclidean distance + epsilon), where epsilon is a small constant to prevent division by zero, or the similarity value calculated using the radial basis function (RBF) kernel function: Similarity = exp(-gamma * Euclidean_Distance^2), where gamma is an adjustable parameter, and the setting of D1 should be adjusted according to the actual data distribution and domain knowledge to balance recall and precision.

[0103] S4: Use a neural network formed by heterogeneous graphs to perform embedding learning on heterogeneous graphs, and generate node vector representations that integrate semantic and structural information;

[0104] After the heterogeneous graph is constructed, the system enters the graph neural network processing stage, inputting the constructed heterogeneous graph into the graph neural network (GNN) model for deep representation learning to generate a low-dimensional embedding vector for each node. The GNN model is a heterogeneous graph neural network based on graph attention network (GAT), which can effectively handle the differences in node and edge types in heterogeneous graphs and adaptively learn the importance of different types of information. The GNN model processing process specifically includes the following steps.

[0105] Step S410: Node-level attention mechanism;

[0106] For each node, the GNN model employs a node-level attention mechanism to calculate the importance of its neighboring nodes in learning the node's own representation. This mechanism takes into account the heterogeneity of different types of neighboring nodes. Specifically, for a central node v and its neighboring node u, the model learns attention coefficients α(v, u). These attention coefficients are determined by the feature vectors of nodes v and u and the type of edges between them. The attention coefficients are calculated through a shared neural network (e.g., a single-layer feedforward network) and normalized using Softmax. This process allows the model to selectively focus on more important neighboring nodes, thereby capturing local structural information. For example, for a commenter node, commenter nodes with edges sharing a common intent may receive higher attention weights than other types of neighboring nodes.

[0107] Step S420: Semantic attention mechanism;

[0108] The GNN model further introduces a semantic-level attention mechanism to evaluate the importance of different meta-paths in aggregating node information. These different meta-paths are specifically paths composed of different types of nodes and edges. For example, for a commenter node U, its information aggregation might originate from the UCI meta-path (connecting to intent I via comment C, reflecting the commenter's intent distribution) and the UU meta-path (connecting to other commenters U, reflecting the commenter's social or collaborative circle). The semantic-level attention mechanism learns the weights of each meta-path through a weight matrix and a bias vector, and normalizes these weights using a softmax function to highlight the meta-paths most effective for identifying troll groups. For instance, if troll groups tend to coordinate under a specific intent, the UCI meta-path might receive higher semantic-level attention weights.

[0109] Step S430: Heterogeneous information aggregation;

[0110] Based on node-level attention and semantic-level attention mechanisms, the GNN model aggregates the information of each node. For a central node v, its new representation H'v is obtained by aggregating the information transmitted along each predefined meta-path to obtain a meta-path-specific node. Aggregation involves summing the neighbor feature vectors after node-level attention weighting, and can be transformed by a linear transformation or a non-linear activation function.

[0111] Step S440: Multi-layer information transmission;

[0112] GNN models stack multiple heterogeneous information aggregation layers to capture deeper graph structure information and complex relationships between nodes. The node embedding vector output by each layer serves as the input to the next layer. After multiple iterations, a low-dimensional embedding vector rich in semantic and structural information for each node is finally generated. Non-linear activation functions and Dropout layers are usually added between layers to prevent overfitting. The resulting node embedding vectors typically have a dimension between 64 and 256, which can represent the complex attributes of nodes and their topological positions in the graph in a high-dimensional space.

[0113] S5: Cluster analysis based on the embedding vectors of commenter nodes to identify abnormal clusters that conform to the characteristics of online troll behavior;

[0114] After obtaining the node embedding vectors generated by the GNN model, the system enters the water gang identification stage. Based on the representation obtained from deep learning, it accurately identifies suspected water gangs, specifically:

[0115] Step S510: Cluster feature extraction;

[0116] For the embedding vector of the commenter node (U), further feature engineering is performed to extract the features required for cluster identification, including but not limited to: Euclidean distance, cosine similarity, local density, and the number and type distribution of edges connecting to other commenter nodes.

[0117] In the above, the Euclidean distance and cosine similarity of the embedding vectors are used to measure the semantic and behavioral similarity of commenters in the embedding space; the local density is calculated by the K-nearest neighbor algorithm, for example, the number of other commenter nodes within a radius R around each commenter node in the embedding space; and the number and type distribution of edges connecting to other commenter nodes, specifically including how many UU common intent edges a commenter connects to, how many UU behavioral pattern edges, and the average weight of these edges. These features comprehensively reflect the commenter's local structure and global position in the graph.

[0118] Step S520: Cluster analysis;

[0119] The system uses the density-based clustering algorithm DBSCAN to cluster the embedding vectors of commenter nodes (U). The DBSCAN algorithm identifies density-connected regions (i.e., clusters) of arbitrary shapes by defining two key parameters: neighborhood radius (Eps) and minimum number of points (MinPts), and marks points that do not belong to any cluster as noise points.

[0120] In this invention, noise points or clusters that are too small are not considered as water gangs. Water gangs usually exhibit synergy and form density clusters of a certain size. The Eps parameter is set empirically based on the k-distance plot to find the inflection point of the curve. MinPts is set according to prior knowledge of the minimum size of water gangs. For example, MinPts is set to 5 to 10 and Eps is set to 0.1 to 0.3.

[0121] Step S530: Identification of abnormal clusters;

[0122] Each cluster identified by the DBSCAN algorithm undergoes feature analysis to determine whether it is a suspected troll group. The criteria include, but are not limited to: the average distance between commenter nodes within the cluster being less than a preset threshold T4; the proportion of comments with the same standardized intent posted by commenter nodes within the cluster being higher than a preset threshold P1; the similarity of behavioral patterns among commenter nodes within the cluster being higher than a preset threshold P2; and an abnormally high number of comments posted by commenter nodes within the cluster within a specific time window, with highly similar or repetitive semantic content. Clusters meeting one or more of the above criteria are marked as suspected troll groups.

[0123] In the above, when the average distance is less than the preset threshold T4, it indicates that the members within the cluster are highly similar in terms of behavioral patterns or intentions. Behavioral patterns include comment posting time and comment length distribution. When the behavioral pattern similarity is higher than the preset threshold P2, it is measured by averaging the behavioral pattern feature vectors of all commenters within the cluster and then calculating the mean cosine similarity between each member and the average vector. If the number of comments posted by commenter nodes within the cluster is abnormally high within a specific time window, and the semantic content of the comments is highly similar or repetitive, specifically, when the average number of comments posted by commenters within the cluster is far greater than 3 standard deviations of the average number of comments posted by the normal user group, and the mean cosine similarity between each pair of their comment texts after BERT embedding exceeds 0.9, it is highly suspected to be a water army activity.

[0124] When a cluster meeting one or more of the above criteria is marked as a suspected troll group, the system generates a report for each suspected group, including a list of its members, a summary of suspicious behavior, and a relevant chain of evidence, including screenshots of similar comments and intent distribution. After troll groups are identified, the system enters a feedback optimization phase. The identified suspected troll groups are fed back to the comment data preprocessing and semantic normalization phases to continuously optimize system performance, forming a dynamically adaptive identification closed loop.

[0125] S6: Feedback the recognition results to the data preprocessing stage to optimize the system's recognition capabilities, specifically:

[0126] The identified suspected troll groups are fed back to the comment data preprocessing and semantic normalization stages to continuously optimize system performance. The feedback optimization process specifically includes optimizing the standardized intent lexicon and reducing the weight of intent statistics.

[0127] In the above, optimizing the standardized intent lexicon involves the system conducting in-depth analysis of the core keyword set associated with the standardized intent used by commenters identified as bots. If bots are found to frequently use certain specific words to disguise legitimate intent (e.g., phrases like "high cost-performance ratio" and "worth buying" being abused in fake positive reviews), these words are marked as high-risk words. These high-risk words are removed from the standardized intent lexicon or their weight is reduced to prevent future legitimate reviews from being misidentified as bot intent due to the inclusion of these words. This weight reduction can be achieved by performing a small multiplicative decay on the embedding vectors of these words in the input word vectors of the intent recognition model. Simultaneously, if the system discovers through semantic clustering or manual review that bots have created new expressions with manipulative intent that are not covered by the current standardized intent lexicon (e.g., new subtle advertising slogans or manipulative rhetoric), the system automatically extracts the key features of these expressions. These features, such as high-frequency co-occurring phrases and phrase structure patterns, are suggested for inclusion in the keyword set of new standardized intents or for updating existing intents, thereby improving the system's adaptability to new bot tactics. This process can be combined with semi-supervised learning, using a small number of manually labeled new types of online troll comments, combined with unlabeled data for iterative learning.

[0128] The statistical weight of intent is reduced. Specifically, for standardized intent nodes highly associated with suspected troll groups, their statistical weight or attention coefficient is dynamically reduced when calculating node embedding vectors in the graph neural network processing module. This means that during subsequent graph neural network learning, the influence of these intents exploited by troll groups on the representations of comment nodes and commenter nodes will be weakened, thereby reducing the effectiveness of troll groups manipulating specific intents to bypass detection. Weight adjustments are achieved through multiplication factors or additive offsets, the magnitude of which is positively correlated with the strength of the association between the intent and the troll group.

[0129] This invention also provides a comment word frequency statistics device based on Internet search. The device is a physical entity or software system architecture that implements the above method, and its functional units work closely together to complete the task of identifying online trolls.

[0130] The device includes a data acquisition unit configured to acquire comment data from an internet platform via a web crawler cluster. The comment data includes the comment text, posting time, commenter identification, identifier of the product / service / event addressed in the comment, and information about the platform from which the comment originated. This unit integrates a crawling rule management module, an IP proxy pool management module, a User-Agent rotation module, and a distributed scheduling module to ensure comprehensive, efficient, and stable data acquisition.

[0131] Furthermore, the device includes a comment data preprocessing and semantic normalization unit, configured to perform preprocessing and semantic normalization operations on the acquired comment data to generate standardized intents. The comment data preprocessing and semantic normalization unit specifically includes:

[0132] The text cleaning and word segmentation module is configured to perform unstructured data cleaning on the comment text, removing HTML tags, special characters, URL links, and emoticons. Subsequently, a word segmentation model based on Conditional Random Field (CRF) is used to perform Chinese word segmentation on the cleaned comment text, obtaining a series of word units, and removing words from a preset stop word set.

[0133] The commenter information extraction and anonymization module is configured to extract commenter identifiers from comment data and perform hash anonymization to generate unique anonymous commenter IDs. Simultaneously, this module extracts metadata related to the commenter, such as the number of historical comments, average comment posting interval, number of followed users, number of followers, and IP address change entropy value, as an initial set of commenter features.

[0134] The intent recognition and normalization module is configured to perform semantic analysis on the cleaned and segmented comment text. By deploying a pre-trained deep learning model based on the Transformer architecture, it maps the input text to a predefined set of normalized intent categories. This module selects the intent with the highest confidence score that is higher than a preset threshold T1 as the primary intent of the comment and normalizes it to the corresponding intent.

[0135] The standardized intent generation module is configured to generate one or more standardized intents for each comment based on the identified standardized intents. An intent includes an intent identifier, intent type, intent confidence level, and a set of core keywords related to that intent extracted from the comment text. The extraction of the core keyword set is achieved by combining part-of-speech tagging results with a text keyword extraction algorithm.

[0136] Furthermore, the device includes a heterogeneous graph construction unit configured to construct a heterogeneous graph G = (V, E) based on preprocessed and semantically normalized data. The heterogeneous graph contains three types of nodes and three types of edges, specifically including:

[0137] The node creation module is configured to create comment nodes (C), commenter nodes (U), and normalized intent nodes (I). This module initializes a feature vector for comment nodes, which is a concatenation of word embeddings of the comment text and comment metadata; initializes a feature vector for commenter nodes, which is a concatenation of commenter metadata; and initializes a feature vector for normalized intent nodes, which is a concatenation of intent embedding vectors and statistical frequencies.

[0138] The edge creation and weight allocation module is configured to create commenter-comment edges (UC edges), comment-intent edges (CI edges), and commenter-commenter association edges (UU edges), and assign weights to each edge. The creation of commenter-commenter association edges (UU edges) is based on shared intent association and behavioral pattern association. Shared intent association is achieved by analyzing the number of comments with the same standardized intent published by the commenter within a preset time window, while behavioral pattern association is achieved by comparing the Euclidean distance of the commenter's behavioral pattern features within the preset time window.

[0139] Furthermore, the device includes a graph neural network processing unit configured to input the constructed heterogeneous graph into a graph neural network (GNN) model for deep representation learning to generate node embedding vectors. The GNN model is a heterogeneous graph neural network based on a graph attention network (GAT) (e.g., HAN). This processing unit integrates:

[0140] The module includes a node-level attention mechanism module for calculating the importance weights of neighboring nodes in learning the representation of the central node; a semantic-level attention mechanism module for evaluating the importance of different meta-paths in aggregating node information; a heterogeneous information aggregation module for aggregating the neighbor information of each node based on the node-level and semantic-level attention mechanisms to generate node embedding vectors rich in semantic and topological information; and a multi-layer information transfer module for stacking multiple heterogeneous information aggregation layers to capture deeper graph structure information and complex relationships between nodes.

[0141] Furthermore, the device includes a troll gang identification unit configured to identify suspected troll gangs based on node embedding vectors generated by a GNN model. The identification process includes:

[0142] The cluster feature extraction module is configured to perform feature engineering on the embedding vector of the commenter node (U) to extract the features required for cluster recognition, such as the Euclidean distance, cosine similarity, local density, and the number and type distribution of edges connecting to other commenter nodes.

[0143] The clustering analysis module is configured to use the density-based DBSCAN algorithm to cluster the embedding vectors of commenter nodes (U) and identify density-connected regions.

[0144] The abnormal cluster identification module is configured to perform feature analysis on each cluster identified by the DBSCAN algorithm to determine whether it is a suspected online troll group. Judgment criteria include the average distance between commenter nodes within the cluster, the proportion of comments posting the same standardized intent, behavioral pattern similarity, and an abnormally high number of comments posted within a specific time window with highly similar or repetitive semantic content.

[0145] Furthermore, the device includes a feedback optimization unit configured to feed back identified suspected troll groups to the comment data preprocessing and semantic normalization stages to continuously optimize system performance. The feedback optimization unit specifically includes:

[0146] The standardized intent lexicon module has been optimized and configured to analyze the standardized intents and related keyword sets used by online troll groups. This module marks specific words used to disguise legitimate intents as high-risk words and removes or reduces their weight from the standardized intent lexicon. Simultaneously, it automatically extracts key features of new, manipulative expressions created by online troll groups and suggests incorporating them into new standardized intents or updating the keyword set of existing intents.

[0147] The intent statistical weight reduction module is configured to dynamically reduce the statistical weight or attention coefficient of standardized intent nodes that are highly associated with suspected troll groups when calculating node embedding vectors in the graph neural network processing module, thereby weakening the effectiveness of troll groups in bypassing detection by manipulating specific intents.

[0148] This invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described method for statistical analysis of comment word frequency based on internet search. The computer-readable storage medium is a non-transitory storage medium, meaning that it can permanently store data and program code, rather than only temporarily storing them during program execution. Examples of storage media include, but are not limited to: hard disk drives (HDDs), solid-state drives (SSDs), optical discs (such as CD-ROMs, DVD-ROMs, and Blu-ray discs), flash memory drives (such as USB flash drives and SD cards), magnetic tapes, and various types of random access memory (RAM, such as DRAM and SRAM). These storage media can store program instructions and data in binary code form. When loaded and executed by one or more processors in a computer system, the processor processes the data according to these instructions, thereby implementing the various functions and steps of this invention, ultimately achieving the purpose of identifying and optimizing online troll groups.

[0149] Example 1: Identifying Paid Product Reviews on E-commerce Platforms

[0150] This system first deployed a web crawler cluster to collect all reviews of the same product from the same e-commerce platform within the past month. During the crawling process, a random delay of 2-10 seconds was set to simulate human behavior, and a proxy pool containing 5,000 User-Agents from commonly used browsers and 20,000 highly anonymous IPs was used to ensure the concealment and stability of the data collection. Ultimately, approximately 1.2 million raw review records were collected, each including a review ID, review text, user ID, posting time, product ID, and source URL.

[0151] Next, the comment data undergoes preprocessing and semantic normalization. During text cleaning, HTML tags, URLs, emojis, and special symbols are removed. A pre-trained CRF segmentation model is then used for word segmentation, filtering out words containing 2500 Chinese stop words. Simultaneously, the system extracts 100,000 unique user IDs, hashes them using the SHA-256 algorithm with a dynamic salt value for anonymization, generating anonymous commenter IDs. Metadata for each commenter is calculated, including the number of historical comments, average comment posting interval, number of followers, number of followings, and IP address change entropy. For intent recognition, a model based on the BERT-Base, Chinese architecture is deployed. This model, fine-tuned on 500,000 labeled corpora, can identify 18 standardized intent categories, with a confidence threshold of 0.8. For example, the above comment is identified as supporting intent with a confidence score of 0.92, generating a standardized intent I-001. Its core keywords are extracted using the TextRank algorithm as {phone, awesome, battery life, powerful capabilities, fast delivery, worth buying}.

[0152] Based on the preprocessed data, the system constructed a heterogeneous graph. The graph contains three types of nodes: comment nodes, commenter nodes, and normalized intent nodes. Comment node features are composed of BERT embedding vectors, comment length, sentiment score, and posting time; commenter node features consist of their metadata; and intent node features are composed of BERT embeddings of intent names and statistical frequencies. Edges include UC edges between commenters and comments, CI edges between comments and intents, and UU edges between commenters. UU edges are created through shared intent associations and behavioral pattern associations. The former is based on the number of comments sharing the same intent within a certain time window, while the latter is calculated based on the Euclidean distance of user behavioral features. Ultimately, the system generated approximately 500,000 shared intent association edges and 300,000 behavioral pattern association edges.

[0153] Subsequently, the heterogeneous graph is fed into a two-layer HAN graph neural network model for processing. This model employs node-level and semantic-level attention mechanisms, aggregating heterogeneous information by learning the importance weights of different neighboring nodes and meta-paths. For example, during training, it was found that the UCI meta-path has a higher weight in identifying fake intent trolls, while the UU meta-path is better at capturing collaborative behavior. The model is trained for 100 epochs using the Adam optimizer, ultimately generating 128-dimensional embedding vectors for all commenter nodes.

[0154] In the bot group identification phase, the system extracts cluster features from the embedding vectors of commenter nodes, including local density and the number of connected UU edges, and uses the DBSCAN algorithm for clustering with parameters set to Eps=0.2 and MinPts=7. After clustering, approximately 1500 clusters were identified, of which 200 were relatively large. By analyzing the average distance between commenter nodes within each cluster, the proportion of comments with the same intent, behavioral pattern similarity, and anomalies in comment quantity, the system identified 45 suspected bot groups. For example, cluster A, containing 12 commenters, posted over 500 comments in two days, 85% of which were highly similar supportive comments, and the standard deviation of their posting intervals was significantly lower than that of normal users, ultimately confirming them as a bot group.

[0155] Finally, the system performs feedback optimization, adjusting the standardized intent word library and weights for the identified online troll behaviors.

[0156] Comparative Example 1: Traditional Online Troll Identification Methods Based on Keywords and Time Windows

[0157] In the data cleaning stage, this comparison method adopted a basic text cleaning process similar to that of this invention. Subsequently, the Jieba word segmentation tool was used for processing, and a keyword library containing 200 common advertising terms and 100 extremely positive review terms was constructed for analysis. Its core identification logic relies on rule matching: keyword matching is performed on each comment and the hit rate is calculated; the total number of comments and the number of duplicate comments for each user within one hour are calculated; and the IP addresses of the comments are recorded for correlation analysis.

[0158] After applying the above rules, this comparison method identified 32 suspected online troll groups or individuals. Among them, 18 groups overlapped with those identified by this invention, indicating some overlap between the two methods in identifying explicit violation patterns. However, this method reveals significant limitations. First, it suffers from severe underreporting of highly disguised online trolls. For example, group A, successfully identified by this invention, had members who deliberately avoided keyword repetition in their comments and whose posting times were relatively dispersed. However, its core characteristics lie in the high consistency of supporting intent and the abnormally concentrated posting intervals. These deep behavioral patterns cannot be captured by traditional rule matching, leading this method to misclassify it as a normal user. Second, this method has a high false positive rate because it relies excessively on keywords and simple thresholds. Finally, this method lacks adaptability; its rule system is fixed and rigid. Once the online troll strategy evolves, such as adopting new vocabulary or adjusting the posting rhythm, the original rules immediately become invalid, unable to be dynamically optimized through a feedback mechanism like this invention.

[0159] Performance comparison data further highlights the differences: the accuracy of the comparison method was 65.2%, the recall was 48.7%, and the F1 score was only 55.8%; while its false positive rate was as high as 12.5%, and its recall rate was significantly lower than the 91.5% of the present invention. These data confirm that its effectiveness is limited in handling complex and disguised coordinated online troll behavior, highlighting the superiority of the graph neural network and deep behavioral pattern analysis method used in the present invention.

[0160] To quantify the performance difference between the methods in Embodiment 1 and Comparative Example 1, we verified the accuracy of the identification results through manual annotation. In the aforementioned new product review dataset, we randomly selected 5000 reviews and their corresponding users, and our expert team manually annotated them to identify genuine paid reviews and paid comment groups. Based on this, we calculated the performance metrics of the two methods, and the results are shown in Table 1.

[0161] Table 1

[0162] Performance indicators Example 1 Comparative Example 1 Recognition accuracy 91.5% 78.2% accuracy 88.7% 72.5% Recall rate 93.2% 68.9% False alarm rate 7.1% 18.3% underreporting rate 6.8% 31.1% Recognition efficiency 12.5s 8.8s

[0163] Comparative analysis of results:

[0164] As can be seen from the data in Table 1, the method of the present invention is significantly superior to the traditional rule-based method of the comparative example in all key performance indicators.

[0165] Identification accuracy: The accuracy of the method of this invention reaches 91.5%, which is much higher than the 78.2% and 70.6% of the comparative methods. This shows that the present invention can identify water gangs more comprehensively and accurately, and its overall performance advantage is obvious.

[0166] Precision and Recall: This invention achieves a precision of 88.7% and a recall of 93.2%. The high recall means that this invention can effectively capture most genuine online troll groups, not just those that are obvious. At the same time, the high precision also indicates that the proportion of genuine online trolls among the identified groups is relatively high, reducing false positives. In contrast, the comparative method lags significantly behind in both precision and recall, especially with a recall rate of only 68.9%, indicating insufficient ability to identify highly disguised online troll groups, and a large number of online trolls going undetected.

[0167] False positive rate and false negative rate: The false positive rate of this invention is only 7.1%, far lower than the 18.3% of the comparative method. This means that the number of normal users being wrongly labeled as shills is greatly reduced, improving user experience and system credibility. Meanwhile, the false negative rate of this invention is also only 6.8%, far lower than the 31.1% of the comparative method. This indicates that this invention is more comprehensive in detecting potential shills and reduces the probability of shills escaping.

[0168] Recognition efficiency: Although the processing time per 10,000 comments is slightly higher than that of the comparative method due to the introduction of complex calculations such as the GNN model, the improvement in recognition accuracy and robustness brought about by this invention is a qualitative leap. Moreover, in a distributed computing environment, this processing time is completely acceptable and can be optimized.

[0169] In summary, this invention, by constructing heterogeneous graphs, utilizing graph neural networks for deep representation learning, and establishing a feedback optimization mechanism, can significantly improve the identification accuracy of highly disguised and collaborative online troll groups, reducing false positive and false negative rates, and providing strong technical support for purifying the internet comment ecosystem. Based on the above-described preferred embodiments of this invention, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification but must be determined according to the scope of the claims.

Claims

1. A method for statistical analysis of comment word frequency based on internet search, characterized in that, Includes the following steps: S1: Obtain comment data from internet platforms, including comment text, comment posting time, publisher ID, factual identifiers of the comments, and information about the platform from which the comments originated; S2: Clean, segment, and identify the intent of the comment text to generate an intent containing standardized intent categories and core keywords, anonymize the commenter's identity, and extract behavioral features; S3: Construct a heterogeneous graph containing comment nodes, commenter nodes, and intent nodes, and associate commenters, comments, and intents; S4: Use a neural network formed by heterogeneous graphs to perform embedding learning on heterogeneous graphs, and generate node vector representations that integrate semantic and structural information; S5: Cluster analysis based on the embedding vectors of commenter nodes to identify abnormal clusters that conform to the characteristics of online troll behavior; S6: Feedback the recognition results to the data preprocessing stage to optimize the system's recognition capabilities.

2. The method for statistical analysis of comment word frequency based on Internet search according to claim 1, characterized in that: In step S2, the specific operation of generating an intent containing standardized intent categories and core keywords is as follows: S21: Perform noise reduction, word segmentation, and stop word removal on the comment text; S22: Hash the commenter's identity and extract their historical behavioral characteristics; S23: Map comment text to standardized intent categories using a pre-trained intent classification model; S24: Extract keywords related to the intent to form a standardized intent.

3. The method for statistical analysis of comment word frequency based on Internet search according to claim 2, characterized in that: The intent classification model is specifically based on the Transformer architecture, and outputs the confidence level of the text belonging to each preset intent category.

4. The method for statistical analysis of comment word frequency based on Internet search according to claim 1, characterized in that: In S3, the specific correspondence is as follows: a commenter node-comment node edge is created based on the comment posting relationship; a comment node-intent node edge is created based on the comment-intent mapping relationship; and a commenter node-comment node edge is created based on the common intent or behavioral pattern similarity between commenters. Commenter nodes and comment node edges represent the comment posting relationship; The edge weights between comment nodes and intent nodes are determined by the confidence level of intent recognition. Commenter nodes and commenter node edges are created and weighted based on the frequency of shared intentions or the similarity of behavioral characteristics among commenters.

5. The method for statistical analysis of comment word frequency based on Internet search according to claim 1, characterized in that: In S4, the heterogeneous graph neural network includes node-level and semantic-level attention mechanisms, supporting information aggregation of multiple types of nodes and edges.

6. The method for statistical analysis of comment word frequency based on Internet search according to claim 1, characterized in that: In S5, a density clustering algorithm is used to identify abnormally dense clusters among commenter nodes, and the consistency of intent and similarity of behavior within the clusters are combined to determine the presence of paid commenters.

7. A comment word frequency statistics device based on Internet search, characterized in that, include: The data acquisition unit is used to acquire comment data on the Internet platform. The comment data includes comment text, comment posting time, commenter identity identifier, identifier targeted by the comment, and information about the platform from which the comment originated. The comment data preprocessing and semantic normalization unit is configured to perform preprocessing and semantic normalization operations on the acquired comment data to generate standardized intents; The heterogeneous graph construction unit is used to construct a heterogeneous graph based on the preprocessed and semantically normalized data. The graph neural network processing unit is used to input the constructed heterogeneous graph into the graph neural network model for deep representation learning in order to generate the embedding vector of the node; The anti-troll group identification unit identifies suspected anti-troll groups based on the node embedding vectors generated by the graph neural network model. as well as The feedback optimization unit is used to feed back the identified suspected troll groups to the comment data preprocessing and semantic normalization stage to continuously optimize system performance.

8. The comment word frequency statistics device based on Internet search according to claim 7, characterized in that: The naval identification unit includes: The feature extraction module is used to extract cluster recognition features from the commenter node embedding vector; The clustering module is used to identify abnormally dense clusters using a density clustering algorithm; The judgment module is used to identify online trolls based on the consistency of intent and similarity of behavior within a cluster.

9. The comment word frequency statistics device based on Internet search according to claim 7, characterized in that: The graph neural network unit includes a heterogeneous graph neural network with node-level and semantic-level attention mechanisms, supporting information aggregation of multiple types of nodes and edges.

10. The comment word frequency statistics device based on Internet search according to claim 7, characterized in that: The intent recognition module is configured as a deep learning model based on the Transformer architecture, and outputs the confidence level of the text belonging to each preset intent category.

Images