Forum public opinion monitoring method and device and electronic equipment

Abstract

The invention discloses a forum public opinion monitoring method and device and electronic equipment. The method comprises the following steps: acquiring forum event data and constructing a target propagation tree; and inputting the target propagation tree into a bidirectional hierarchical semantic interaction neural network model to obtain a credibility classification result of the forum event, thereby realizing forum public opinion monitoring. The bidirectional hierarchical semantic interaction neural network model is a deep learning model for outputting forum public opinion event credibility classification based on a tree propagation structure through an interaction mechanism of top-to-bottom multi-branch semantic coding and bottom-to-top global semantic representation in combination with transverse semantic feature extraction between same-level nodes. According to the method, multi-dimensional semantic information of a propagation tree can be accurately covered, and semantic features of information propagation can be comprehensively reflected, so that the core capability of public opinion monitoring is remarkably improved, a high-precision credibility classification result is output, the capability of identifying abnormal public opinions and normal public opinions is enhanced, and the accuracy and reliability of monitoring are practically improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method, device, and electronic equipment for monitoring forum public opinion. Background Technology

[0002] Forum public opinion monitoring is of great significance for timely identification of abnormal public opinion and control of public opinion trends. Its core requirement is to accurately capture multi-source semantic information in the dissemination process.

[0003] Existing RvNN-based public opinion monitoring methods (such as bottom-up BU models, etc.)

[0004] The top-down TD model has obvious limitations: On the one hand, this type of method only focuses on the vertical propagation path of "top-down" or "bottom-up", and can only capture the hierarchical semantics of comments and reposts, but ignores the horizontal semantic relationships between nodes at the same level in the tree structure (such as the echoing, supplementing or refuting of opinions between comments at the same level), resulting in the loss of the breadth semantic information of the propagation tree and the inability to use multi-source information at the same level to support public opinion judgment; on the other hand, the existing RvNN architecture lacks a dedicated semantic fusion encoder, and can only merge node representations through simple recursion, which cannot efficiently integrate "vertical propagation features, multi-branch semantics, and peer interaction information", resulting in fusion gaps in multi-source semantics, which ultimately affects the accuracy and robustness of public opinion recognition. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the aforementioned shortcomings of existing technologies by proposing a method, apparatus, and electronic device for monitoring forum public opinion. This method can improve the accuracy and robustness of forum public opinion identification.

[0006] In a first aspect, the present invention provides a forum sentiment monitoring method based on a bidirectional hierarchical semantic interaction neural network, the method comprising the following steps:

[0007] Acquire forum event data and construct a target propagation tree;

[0008] The target propagation tree is input into a bidirectional hierarchical semantic interaction neural network model to obtain the credibility classification results of forum events, thereby realizing forum public opinion monitoring.

[0009] Among them, the bidirectional hierarchical semantic interaction neural network model is a deep learning model that outputs the credibility classification of forum public opinion events based on a tree-like propagation structure. It uses an interaction mechanism of top-down multi-branch semantic encoding and bottom-up global semantic representation, combined with the extraction of horizontal semantic features between peer nodes.

[0010] Furthermore, before inputting the target propagation tree into the bidirectional hierarchical semantic interaction neural network model, the method also includes: constructing a bidirectional hierarchical semantic interaction neural network model;

[0011] Constructing a bidirectional hierarchical semantic interaction neural network model specifically includes:

[0012] Step A1: Collect historical tree-like propagation events and perform structured tree processing on the historical tree-like propagation events to obtain tree-like propagation events; and perform text preprocessing and abnormal public opinion labeling on the tree-like propagation events to form a tree-like sample set;

[0013] Step A2: Divide the tree-structured sample set into a training set and a test set according to a preset ratio;

[0014] Step A3: Based on the training set, iteratively train the initial semantic interaction neural network model to obtain the initial classification and recognition model;

[0015] Step A4: Input the test set into the initial classification and recognition model to obtain the test classification result for each test sample in the test set;

[0016] Step A5: Based on the test classification results, adjust the parameters of the initial semantic interaction neural network model to obtain a bidirectional hierarchical semantic interaction neural network model.

[0017] Further, step A3 specifically includes:

[0018] Step B1: Set the initial parameters of the initial semantic interaction neural network model; the initial parameters include batch size, stopping condition, and maximum number of iterations;

[0019] Step B2: Input the training set into the initial semantic interaction neural network model in batches according to batch size;

[0020] Step B3: During the initial semantic interaction neural network model operation, based on the tree-like features, generate independent classification results for each training sample in the training set;

[0021] Step B4: Based on the classification results, update the model parameters of the initial semantic interaction neural network model and output the updated model parameters;

[0022] Step B5: Repeat steps B2 to B4 until the maximum number of iterations is reached or the stopping condition is met;

[0023] Step B6: Based on the updated model parameters of the final output, determine the initial semantic interaction neural network model as the initial interaction recognition model.

[0024] Furthermore, prior to step A3, the method also includes the step of: constructing an initial semantic interaction neural network model;

[0025] Constructing an initial semantic interaction neural network model specifically includes:

[0026] Construct a propagation tree; where the parent node of the propagation tree is the original post, and the child nodes of the propagation tree are the comments;

[0027] Encode both parent and child nodes into sentence vectors;

[0028] Based on the sentence vector, breadth information is extracted from comments at the same level to obtain a breadth information vector; based on the sentence vector, semantics are passed down from top to bottom and the parent node influences the child node to obtain a top-down information feature vector; based on the sentence vector, information is aggregated from bottom to top, and all child node information is summarized to the corresponding parent node to obtain a bottom-up feature vector.

[0029] The breadth information vector, the top-down information feature vector, and the bottom-up feature vector are concatenated to obtain a multi-feature composite vector.

[0030] The multi-feature integrated vector is input into a pre-selected classification learning model to obtain an initial semantic interaction neural network model.

[0031] Furthermore, the breadth information vector is extracted using multiple convolutional kernels.

[0032] Furthermore, based on the sentence vector, semantics are passed down from top to bottom, and parent nodes influence child nodes to obtain a top-down information feature vector, specifically including:

[0033] Determine the top-down transmission path;

[0034] Based on the top-down transmission path, the sentence vector of the parent node and the sentence vector of the corresponding child node are semantically fused to obtain the first semantic information;

[0035] If a parent node corresponds to multiple child node branches, then an attention mechanism is used to adjust and filter the semantic information of the sentence vector of the parent node and the sentence vector of the corresponding multi-branch child nodes.

[0036] By summarizing the first semantic information, a top-down information feature vector is obtained;

[0037] Based on the sentence vector, information is aggregated from bottom to top, summarizing all child node information to the corresponding parent node to obtain a bottom-up feature vector, which specifically includes:

[0038] Determine the bottom-up transmission path;

[0039] Based on the bottom-up transmission path, the sentence vectors of all child nodes under the same parent node are summarized and refined level by level to obtain the comprehensive semantic vector of the child nodes.

[0040] The sentence vector of the parent node is semantically fused with the comprehensive semantic vector to obtain the second semantic information;

[0041] By summarizing the second semantic information, a bottom-up feature vector is obtained.

[0042] Furthermore, the breadth-based information vector, the top-down information feature vector, and the bottom-up feature vector are concatenated to obtain a multi-feature composite vector, which specifically includes:

[0043] The top-down information feature vector and the bottom-up feature vector are concatenated to obtain the depth information vector;

[0044] By concatenating the breadth information vector and the depth information vector, a multi-feature composite vector is obtained.

[0045] Secondly, the present invention provides a forum public opinion monitoring device based on a bidirectional hierarchical semantic interaction neural network, the device comprising:

[0046] The acquisition unit is used to acquire forum event data and construct a target propagation tree;

[0047] The input unit, connected to the acquisition unit, is used to input the target propagation tree into the bidirectional hierarchical semantic interaction neural network model to obtain the credibility classification results of forum events, thereby realizing forum public opinion monitoring.

[0048] Among them, the bidirectional hierarchical semantic interaction neural network model is a deep learning model that outputs the credibility classification of forum public opinion events based on a tree-like propagation structure. It uses an interaction mechanism of top-down multi-branch semantic encoding and bottom-up global semantic representation, combined with the extraction of horizontal semantic features between peer nodes.

[0049] Furthermore, the device also includes building blocks;

[0050] The building unit is connected to the input unit to construct a bidirectional hierarchical semantic interaction neural network model;

[0051] The building blocks specifically include:

[0052] The preprocessing module is used to collect historical tree-like propagation events and perform structured tree-like processing on these events to obtain tree-like propagation events; and to perform text preprocessing and abnormal public opinion labeling on the tree-like propagation events to form a tree-like sample set.

[0053] The partitioning module, connected to the preprocessing unit, is used to divide the tree-like sample set into training set and test set according to a preset ratio.

[0054] The training module, connected to the partitioning module, is used to iteratively train the initial semantic interaction neural network model based on the training set to obtain the initial classification and recognition model.

[0055] The testing module, connected to the training module, is used to input the test set into the initial classification and recognition model and obtain the test classification result for each test sample in the test set.

[0056] The adjustment module, connected to the testing module, is used to adjust the parameters of the initial semantic interaction neural network model based on the test classification results, thereby obtaining a bidirectional hierarchical semantic interaction neural network model.

[0057] Thirdly, the present invention provides an electronic device including a memory and a processor. The memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the forum sentiment monitoring method based on a bidirectional hierarchical semantic interaction neural network as described in the first aspect.

[0058] This invention constructs a target propagation tree by acquiring forum event data and inputs it into a bidirectional hierarchical semantic interaction neural network model to output credibility classification results, thereby accurately achieving forum public opinion monitoring. The specific beneficial effects are as follows:

[0059] (1) Accurate coverage of the semantic dimension of the propagation tree: The bidirectional hierarchical semantic interaction neural network model is designed based on the tree propagation structure. It integrates the interaction mechanism of "top-down multi-branch semantic encoding" and "bottom-up global semantic representation" and the function of "horizontal semantic feature extraction between peer nodes". It can capture the vertical propagation semantics of comments, reposts and other hierarchical messages, and also obtain the horizontal semantic associations between peer nodes (such as echoing and supplementing viewpoints). It can effectively avoid the information loss caused by a single semantic dimension and provide more comprehensive semantic support for public opinion judgment.

[0060] (2) Enhance core capabilities of public opinion monitoring: By combining a two-way hierarchical semantic interaction mechanism with horizontal feature extraction, the two-way hierarchical semantic interaction neural network model can effectively process the multi-dimensional semantics of the propagation tree and finally output accurate forum event credibility classification results. Compared with methods that rely solely on vertical propagation paths, it can more accurately identify public opinion attributes (such as abnormal public opinion and normal public opinion) and effectively improve the accuracy and reliability of forum public opinion monitoring. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the forum public opinion monitoring method based on a bidirectional hierarchical semantic interaction neural network in an embodiment of the present invention;

[0062] Figure 2 This is a diagram illustrating the overall framework of forum sentiment monitoring based on a bidirectional hierarchical semantic interaction neural network in this invention embodiment.

[0063] Figure 3 This is a schematic diagram of the tree structure in an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of the encoding based on a bidirectional hierarchical semantic interaction neural network in an embodiment of the present invention;

[0065] Figure 5 This is a schematic diagram of a multi-branch propagation chain based on a bidirectional hierarchical semantic interaction neural network in an embodiment of the present invention;

[0066] Figure 6 This is a schematic diagram of a forum public opinion monitoring device based on a bidirectional hierarchical semantic interaction neural network in an embodiment of the present invention.

[0067] Figure 7 This is an architectural diagram of an electronic device according to an embodiment of the present invention.

[0068] Reference numerals: 10, acquisition unit; 20, input unit; 100, processor; 200, memory. Detailed Implementation

[0069] To enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0070] It is understood that the specific embodiments and accompanying drawings described herein are merely for explaining the invention and are not intended to limit the invention.

[0071] It is understood that, without conflict, the various embodiments and features in the embodiments of the present invention can be combined with each other.

[0072] It is understood that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, while the parts unrelated to the present invention are not shown in the drawings.

[0073] It is understood that each unit or module involved in the embodiments of the present invention may correspond to only one entity structure, or may be composed of multiple entity structures, or multiple units or modules may be integrated into one entity structure.

[0074] It is understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of this invention may occur in a different order than that marked in the accompanying drawings.

[0075] It is understood that the flowcharts and block diagrams of this invention illustrate the possible architecture, functions, and operations of systems, apparatuses, devices, and methods according to various embodiments of this invention. Each block in the flowchart or block diagram may represent a unit, module, program segment, or code, containing executable instructions for implementing the specified function. Furthermore, each block or combination of blocks in the block diagram and flowchart can be implemented using a hardware-based system to achieve the specified function, or using a combination of hardware and computer instructions.

[0076] It is understood that the units and modules involved in the embodiments of the present invention can be implemented by software or by hardware. For example, the units and modules can be located in a processor.

[0077] The rapid spread of abnormal public opinion and false information on social media and online forums has far-reaching harms. It can not only disrupt social stability (e.g., misinformation causing public panic), but also affect economic order (e.g., false financial information causing stock market fluctuations), and even damage individual mental health, triggering anxiety, panic, and a crisis of social trust. Therefore, conducting efficient and accurate public opinion monitoring is particularly important. Current public opinion detection methods largely revolve around the propagation tree of abnormal public opinion, resulting in various technical approaches, including schemes using GRU as the core combined with tree structures and RNNs, frameworks integrating RNNs and CNNs, and multivariate schemes that integrate user information and temporal features. Among these, recurrent neural networks (RvNNs) are one of the core models. From the early matrix-vector recurrent neural network (MV-RNN) to the subsequent bottom-up (BU) and top-down (TD) variants, all refine node features through recursion to improve detection efficiency. However, existing technologies still have significant shortcomings: they struggle to fully capture the broad semantic information at the same level within the propagation tree, cannot fully utilize top-down multi-branch semantics through weight adjustment, and lack an effective encoder to summarize the key semantics in the "comment + repost" propagation chain, resulting in insufficient utilization of propagation information. To address these issues, this invention proposes a forum sentiment monitoring scheme: first, a sentiment propagation tree is constructed based on event reposts and comments; then, a CNN encoder and multiple convolutional kernels are used to capture the broad semantic encoding at the same level; simultaneously, a bottom-up semantic synthesis representation is combined, and an attention mechanism is used to adjust the semantic weights of each branch from top to bottom to filter key information; finally, the key clues from multiple branches and the broad semantics are fused. This scheme includes three embodiments, covering the monitoring method in Embodiment 1, the device in Embodiment 2, and the electronic device in Embodiment 3.

[0078] Example 1:

[0079] This embodiment provides a forum public opinion monitoring method based on a bidirectional hierarchical semantic interactive neural network. This method can be applied to various online forum scenarios, such as monitoring public opinion related to hot events in news forums, and can promptly identify whether there is the spread of abnormal public opinion or inappropriate remarks. It is also applicable to the exchanges between users in social forums around social phenomena and public topics, helping platforms or regulators to quickly grasp the trend of public opinion, prevent the spread of false information and negative public opinion, and provide technical support for maintaining the order of cyberspace and ensuring that the public can obtain true information.

[0080] like Figure 1As shown, the forum public opinion monitoring method based on bidirectional hierarchical semantic interaction neural network includes the following steps S1 to S2.

[0081] Step S1: Obtain forum event data and construct the target propagation tree.

[0082] Specifically, the process of acquiring forum event data and constructing a target propagation tree first requires extracting relevant comment data from the database, including user information, comment content, parent nodes, and current node numbers. This structured data is then preprocessed and organized, and a tree structure is built according to the hierarchical relationships between comments. For example, using "former_index" as the parent node number, each comment is connected to its corresponding parent comment, constructing the hierarchical relationship of the tree layer by layer. Ultimately, the resulting propagation tree clearly displays the propagation path and hierarchical structure of comments, which helps analyze information propagation patterns, identify abnormal public opinion propagation chains, and thus achieve effective monitoring and management of events within the forum.

[0083] Step S2: Input the target propagation tree into the bidirectional hierarchical semantic interaction neural network model to obtain the credibility classification results of forum events, thereby realizing forum public opinion monitoring.

[0084] Specifically, the bidirectional hierarchical semantic interaction neural network model is a deep learning model that outputs the credibility classification of forum public opinion events based on a tree-like propagation structure. It uses an interaction mechanism of top-down multi-branch semantic encoding and bottom-up global semantic representation, combined with the extraction of horizontal semantic features between peer nodes.

[0085] As a specific implementation method, a bidirectional hierarchical semantic interaction neural network model is constructed, specifically including steps A1 to A5.

[0086] Step A1: Collect historical tree-like propagation events and perform structured tree processing on the historical tree-like propagation events to obtain tree-like propagation events; and perform text preprocessing and abnormal public opinion labeling on the tree-like propagation events to form a tree-like sample set.

[0087] Step A2: Divide the tree-structured sample set into a training set and a test set according to a preset ratio.

[0088] Step A3: Based on the training set, iteratively train the initial semantic interaction neural network model to obtain the initial classification and recognition model.

[0089] Step A3 specifically includes steps B1 to B6.

[0090] Step B1: Set the initial parameters of the initial semantic interaction neural network model; the initial parameters include batch size, stopping condition, and maximum number of iterations;

[0091] Step B2: Input the training set into the initial semantic interaction neural network model in batches according to batch size;

[0092] Step B3: During the initial semantic interaction neural network model operation, based on the tree-like features, generate independent classification results for each training sample in the training set;

[0093] Step B4: Based on the classification results, update the model parameters of the initial semantic interaction neural network model and output the updated model parameters;

[0094] Step B5: Repeat steps B2 to B4 until the maximum number of iterations is reached or the stopping condition is met;

[0095] Step B6: Based on the updated model parameters of the final output, determine the initial semantic interaction neural network model as the initial interaction recognition model.

[0096] Before step A3, the method also includes the step of: constructing an initial semantic interaction neural network model;

[0097] Constructing an initial semantic interaction neural network model specifically includes:

[0098] Construct a propagation tree; where the parent node of the propagation tree is the original post, and the child nodes of the propagation tree are the comments;

[0099] Encode both parent and child nodes into sentence vectors;

[0100] Based on the sentence vector, breadth information is extracted from comments at the same level to obtain a breadth information vector; based on the sentence vector, semantics are passed down from top to bottom and the parent node influences the child node to obtain a top-down information feature vector; based on the sentence vector, information is aggregated from bottom to top, and all child node information is summarized to the corresponding parent node to obtain a bottom-up feature vector.

[0101] The breadth information vector, the top-down information feature vector, and the bottom-up feature vector are concatenated to obtain a multi-feature composite vector.

[0102] The multi-feature integrated vector is input into a pre-selected classification learning model to obtain an initial semantic interaction neural network model.

[0103] As a specific implementation method, the breadth information vector is extracted using multiple convolutional kernels.

[0104] As a specific implementation method, top-down semantic transmission is achieved through a top-down semantic encoding path, and attention mechanisms are used to adjust and filter multi-branch semantic information; bottom-up information aggregation is achieved through a bottom-up semantic aggregation path, and information fusion mechanisms are used to summarize and refine the local semantics of child nodes level by level.

[0105] Specifically, based on the sentence vector, semantics are passed down from top to bottom, and parent nodes influence child nodes to obtain a top-down information feature vector, which includes:

[0106] Determine the top-down transmission path; based on the top-down transmission path, semantically fuse the sentence vector of the parent node with the sentence vector of the corresponding child node to obtain the first semantic information; if the parent node corresponds to multiple child node branches, then use the attention mechanism to adjust and filter the semantic information of the sentence vector of the parent node and the sentence vector of the corresponding multi-branch child nodes; summarize the first semantic information to obtain the top-down information feature vector.

[0107] Based on the sentence vectors, a bottom-up aggregation method is used to summarize all child node information to the corresponding parent node, resulting in a bottom-up feature vector. Specifically, this includes: determining the bottom-up transmission path; summarizing and refining the sentence vectors of all child nodes under the same parent node according to the bottom-up transmission path to obtain the comprehensive semantic vector of the child node; semantically fusing the sentence vector of the parent node with the comprehensive semantic vector to obtain the second semantic information; and summarizing the second semantic information to obtain the bottom-up feature vector.

[0108] As a specific implementation method, the breadth information vector, top-down information feature vector, and bottom-up feature vector are concatenated to obtain a multi-feature composite vector, specifically including:

[0109] The top-down information feature vector and the bottom-up feature vector are concatenated to obtain the depth information vector;

[0110] By concatenating the breadth information vector and the depth information vector, a multi-feature composite vector is obtained.

[0111] The process of obtaining a multi-feature composite vector is illustrated with a specific example:

[0112] There are n events obtained from the forum, and each tree diagram represents one event. Each event is classified into four categories according to its credibility.

[0113] In the classification process, in order to effectively capture semantic information in the abnormal dissemination of public opinion, this invention considers CNN to obtain broad semantic information, and at the same time obtains a comprehensive representation of top-down multi-branch semantic encoding. In order to adjust the weight of multi-branch semantic encoding, effectively utilize the bottom-up comprehensive representation, obtain key clues based on the attention mechanism, and finally obtain good public opinion detection results by combining the semantics in the deep dissemination process with the broad semantics.

[0114] During the widespread dissemination of abnormal public opinion, multi-dimensional deep semantic features are obtained using multiple convolutional kernels:

[0115] X P =Conv(Wr+b)

[0116] X m,o =Conv(Wx m,o +b)

[0117] Where P represents the root node vector, X P x is the result of the root node vector passing through the convolutional layer. m,o Let X be the vector representation of the nodes on the propagation path, where m represents the level of the abnormal public opinion propagation tree, m∈(2,D), D is the depth of the abnormal public opinion propagation tree, o represents the o-th node among sibling nodes, o∈(1,Wid), and Wid represents the number of nodes at the current depth. The final convolutional vector representation is X. m,o W represents the convolution kernel, used to capture key semantic information in each comment, and b is the offset.

[0118] The method described in the patent is a method for detecting abnormal public opinion using a bidirectional hierarchical semantic interaction neural network. Its core feature lies in the following: after each node obtains a new sentence-level feature representation based on TextCNN, information interaction is implemented between nodes within the layer. This process is achieved through one-dimensional TextCNN encoding, where the size of the convolutional kernel W∈R h×d , where h represents the height of the convolution kernel and d is the length of the sentence-level feature representation generated in the previous step. By convolving the last node of the current layer with the first node of the next layer, this method maintains its ability to extract a wide range of semantics while establishing interaction between layers. X i Let X be a matrix composed of nodes in the i-th layer. i,D:i+1,1 The matrix consisting of the last node of layer i and the first node of layer i+1 is subjected to multiple iterations and concatenation operations. Finally, a single convolutional kernel obtains the output vector Z, which is the semantic encoding of breadth information. out1 The Z 0ut1 ∈R 1×s-h+1 s represents the total number of points for this identification event, and the length of this total number of nodes is determined by the number of comments for each event. The specific algorithm flow is as follows:

[0119] Z i =Conv(WX i +b)

[0120] Z i:i+1 =Conv(WX i,D:i+1,1 +b)

[0121] Z i+1 =Conv(WX i+1 +b)

[0122]

[0123] This embodiment is an abnormal public opinion detection technique based on a bidirectional hierarchical semantic interaction neural network. It captures rich inter-sentence interaction features by selecting multiple convolutional kernels. The output vector of the last convolutional kernel is represented as Z. outn , where n represents the number of convolutional kernels, and the resulting multi-layer semantic information is represented by Z. Where Z∈R n×1 ,

[0124]

[0125] In the top-down method for acquiring multi-branch semantic information, the node is represented as a sentence-level vector composed of the TF-IDF value word embeddings of the vocabulary words, j represents the current node, the parent node of the current node j is represented as p(j), and the vector encoded by the parent node p(j) is used as the hidden layer vector h of the current node j. p(j) The input vector for the current node j is the encoding x of the comment information of user j. j The final output vector is h j The specific algorithm is shown below.

[0126]

[0127] Where E is the trainable parameter matrix, E∈R d×5000 d represents the dimension of the hidden layer, 5000 represents the total length of the dictionary, and x j ∈R 1×m m represents the vector x j The length of E is such that, during computation, the dimension of E needs to be determined based on x. j The index subscript is adjusted to...

[0128] E∈R d×m ,final Represented as Among them, W r U r W z U z W h U h ∈R d×d The final output vector h j ∈R 1×d And in matrix form This represents the depth propagation threads of all leaf nodes, where n represents the total number of threads.

[0129] In bottom-up methods for obtaining global comprehensive semantics:

[0130]

[0131] Where, x j Let E represent the original input vector of node j, and let E represent the parameter matrix used to transform the input post. Let h be the transformed representation of node j, where [W,U] represents the weighted connections within the GRU, and h is the weighted connection. j and h s Let h represent the hidden states of node j and its S-th child node, respectively. S This represents the sum of the hidden states of all child nodes of node j, assuming all child nodes are equally important to node j. Similar to the standard GRU, the symbol ⊙ denotes element-wise multiplication; the reset gate r... j Used to determine how to process the current input. Combined with the memory of child nodes, the gate z is updated. j This defines the extent to which the memory of the child node is passed to the current node; This represents the candidate activation value of the current node's hidden state. Unlike standard GRU units, the gating vectors in our GRU variant depend on the states of multiple child units, allowing the model to combine representations from different child nodes.

[0132] After obtaining the top-down and bottom-up representations, this patent performs attention adjustment on the top-down multi-branch semantics using the comprehensive node representation obtained from the bottom-up approach, that is, adjusts the weights of different branches. The specific method is as follows:

[0133]

[0134] u=h M E

[0135]

[0136] Finally, through vertical and horizontal semantic interaction, the final result is obtained, namely the multi-feature comprehensive vector M:

[0137] M = Concat(P, Z).

[0138] M is fed into a multilayer perceptron to obtain the final discrimination result.

[0139] Step A4: Input the test set into the initial classification and recognition model to obtain the test classification result for each test sample in the test set.

[0140] Step A5: Based on the test classification results, adjust the parameters of the initial semantic interaction neural network model to obtain a bidirectional hierarchical semantic interaction neural network model.

[0141] Figure 2This paper demonstrates a deep learning model architecture for processing forum comment propagation trees, combining convolutional neural networks (CNNs) and gated recurrent units (GRUs), and finally performing predictions through an attention mechanism and fully connected layers. This embodiment divides the graph into several main parts: the first part is the input, i.e., the forum comment propagation tree, represented by nodes x1 to x6 and the root node r, simulating the propagation relationship of comments (such as main post, reply, reply to reply, etc.), with each node representing a comment content (text); the second part is branch 1—the CNN branch, responsible for extracting local features or text sequence features of the nodes. This is achieved by inputting node representations (such as word embeddings) into multiple convolutional kernels, extracting features layer by layer, applying pooling (MaxPooling or average pooling) to reduce dimensionality and extract important information, and finally outputting a feature vector Z representing the node. The first part is the local features of the tree; the second part is branch 2—the GRU branch, which processes the sequence relationships in the tree structure. It uses multi-layer GRU to capture the order information and long-term dependencies between comments. The input is the node representation, and the output is the sequence features of the tree structure. The hidden state or intermediate vector u in the processing is used for information transmission and attention calculation; the third part is the fusion part, which combines the features of CNN and GRU (Out1, Out2, ...), uses the attention mechanism to assign weights (α1, α2, ...) to each feature, calculates the attention weights based on the u vector, and fuses the different features into the final feature P, which is then input into the fully connected layer to generate the final prediction result. In the final overall process, the input is a text representation of tree nodes. Local features are extracted using a CNN, sequence relationships are captured using a GRU, multi-feature fusion is performed using an attention mechanism, and the final prediction is completed through a fully connected layer. This architecture fully combines the local feature extraction capabilities of CNNs, the sequence modeling capabilities of GRUs, and the key feature weighting advantages of the attention mechanism. It is specifically designed for analyzing tree-structured comment data and supports various tasks such as sentiment analysis and abnormal public opinion detection.

[0142] To illustrate this with a specific example, the following steps are included:

[0143] (1) On the forum platform, users can freely express their opinions, and the comments are stored in a structured database. Due to the large number of comments, administrators need to filter out negative comments and abnormal public opinion for review, which significantly increases labor costs. To effectively reduce the burden of manual screening, such as... Figure 3As shown, the data is stored in a structured manner, and an innovative tree-structured abnormal public opinion detection model is introduced, into which the structured data is fed. According to the database table structure, there is a table named "data_table" containing the following columns: "User" (recording users who posted comments, such as "User 1", "User 2", etc.), "former_index" (representing the parent node, such as "0" for "User 2" and "1" for "User 5"), "index" (representing the current node, such as "0" for "User 1" and "1" for "User 2"), "Text" (storing the comment content, such as "XXXX" for "User 1"), and "Event" (the event code for all records is "1"). After data preprocessing, the final tree structure is formed. As seen in the tree diagram on the right, the root node is "0", with three child nodes "1", "2", and "3"; node "1" has a child node "4", and node "2" has a child node "5". This tree-like structure can clearly show the hierarchical relationships and connections between comments. By using an innovative tree-like abnormal public opinion detection model to analyze this structured data, it is possible to detect abnormal public opinion and other harmful information more efficiently, reduce the pressure of manual review, and thus reduce labor costs.

[0144] After data processing, the tree-structured data is first represented as text and then fed into the model encoder. In this embodiment, TF-IDF is used to represent the text content, and the represented text is fed into the encoder to obtain rich contextual semantic interaction.

[0145] (2) Figure 4As shown, the CNN (Convolutional Neural Network) encoder's processing of the event propagation tree consists of two core steps: "intra-node convolution" and "inter-node convolution." These two steps complement each other, jointly extracting key semantic features for public opinion detection: 1. Intra-node convolution: Capturing the semantics within a single node. Taking the root node r (number 0) of the propagation tree as an example, it carries a specific event description such as "on a certain day, a certain type of activity and a certain type of problem occurred." By performing convolution operations within a single node, the contextual dependencies and semantic relationships within the node's text can be fully captured—such as fine-grained information like the time, location, activity type, and problem manifestation of the event. The core value of this step lies in extracting a precise semantic representation of "what" and "what happened" for each node from its own text content. 2. Inter-node convolution: Modeling the propagation relationships between nodes. In the event propagation tree, child nodes (such as nodes 1, 2, and 3 with node 0 as their parent, node 4 with node 1 as its parent, and node 5 with node 2 as its parent) are essentially generated "based on the comments or reposts of the previous node," exhibiting strong correlations: node 1 may supplement, refute, or extend node 0; node 4 may further discuss node 1, and so on. To capture these semantic connections between "parent-child," "grandparent-grandchild," and even cross-level nodes, the encoder introduces convolutions (CNNs) between nodes. As can be seen from the range covered by the "kernel" in the diagram (such as combinations of 0 and 1, 0 and 2, 1 and 4, 2 and 5, etc.), it simultaneously "sees" the context of multiple related nodes, learning how these nodes influence each other and how semantics are transmitted or evolved during propagation. In this way, the model can not only understand "what a single node said," but also understand "why different nodes propagate in this way and how they respond to each other," thereby extracting more propagation-oriented semantic features. (Note: the root node is the initial parent node.) 3. Overall Value: Provides key features for public opinion monitoring. After two layers of convolution ("intra-node + inter-node"), the semantic features output by the encoder contain both fine-grained context of individual nodes and propagation correlation information between nodes. These features are crucial for subsequent forum public opinion tasks (such as abnormal public opinion detection, popularity prediction, sentiment analysis, etc.): for "abnormal public opinion detection," accurate semantics within nodes can help determine whether a single piece of content is reasonable, while the correlation between nodes can assist in analyzing "how abnormal public opinion is distorted and spread during propagation." In this embodiment, the CNN encoder transforms the event propagation tree into a semantic representation that is more easily understood by downstream tasks by adopting the approach of "first splitting nodes, then connecting them in series for propagation," allowing the public opinion monitoring model to make more accurate judgments from the dual perspectives of "single node meaning + propagation logic."

[0146] (3) Since the propagation chain of an event is not a single branch but a multi-branch chain structure, an RNN (GRU, i.e., gated recurrent unit, was chosen in this paper) was used to capture the semantics of different propagation branches in order to capture the semantics of this multi-branch chain propagation structure. The core idea is to comprehensively analyze the representations of different propagation chains based on the top-down propagation structure and obtain the feature output of each branch. At the same time, for the global information traced from the bottom up, a complete semantic context can be obtained when all branches converge to the root node. This global semantics serves as a key feature, and the multi-branch nodes are combined with this global feature. The importance and weight of different branches are calculated through an attention mechanism to obtain the weighted representation of each branch, which serves as the final output of the RNN (GRU) channel. Figure 5The model in the example demonstrates a multi-branch propagation chain structure based on an RNN kernel. Specifically, the propagation chain is "multi-branch chain-like," meaning that after a piece of information originates from the root node, it branches into multiple sub-chains. For example, in social media, a post is forwarded and commented on multiple times, forming multiple propagation branches. The purpose of the RNN (GRU) here is to "acquire the semantics of different branches," because each branch has different propagation patterns, participating nodes, and interaction methods. The RNN can process nodes sequentially (by propagation time or logical order), learning the temporal and content features within each branch. "Top-down" here refers to starting from the "root node" of propagation and processing downwards along each branch, during which the RNN learns the local features (branch representations) of each branch. "Bottom-up" refers to integrating the global information of these branches when multiple branch information finally converges at the "root node," serving as "key semantic features." Attention mechanisms play a crucial role here. They combine the encoded global semantics with multi-branch node information to calculate attention weights, which measure the importance of each branch to the final result—important branches receive higher weights, while less important ones receive lower weights. Finally, the weighted combination yields the output of the RNN channels. Taking a tree structure as an example, the two GRU modules on the left represent the processing of different branches, corresponding to the multi-branch topology: the root node 0 branches downwards into three branches: the first is 0→1→4, the second is 0→2→5, and the third is 0→3. Nodes 4 and 5 are leaf nodes, where the propagation chain ends. Each GRU sub-module corresponds to a sub-chain originating from the root, processing nodes sequentially on its chain (e.g., 0→1→4, 0→2→5), learning the temporal features within each branch, and finally generating the representation of that branch (Output01 and Output02). The multi-branch chain-like propagation structure corresponds to the multiple propagation chains in the diagram, such as 0→1→4, 0→2→5, 0→3, etc. The purpose of using GRU is to "acquire the semantics of different branches," meaning each GRU submodule processes one branch independently, learning its features (Output01 and Output02 are the branch representations). Simultaneously, through both "top-down" and "bottom-up" approaches, propagating from the tree's root to the leaves and summarizing from the leaves to the root, the propagation semantics of multiple branches are comprehensively captured. Using an attention mechanism, the Output of each branch (e.g., Output01 and Output02) is combined with the "global semantics" to calculate the weighting coefficient of each branch, determining its contribution to the final prediction, thus achieving hierarchical representation learning of the multi-branch propagation chain. The overall architecture, composed of the tree's topology, GRU submodules for each branch, and global information combined with the attention mechanism, fully covers the entire process of "semantic capture and integration of the multi-branch propagation chain."

[0147] (4) Finally, the breadth and depth semantic information encoded by TextCNN and RNN are input into the final fully connected layer to obtain the final detection result. This case clearly demonstrates the details from data extraction and structured processing to semantic representation within the model. This patent enables efficient screening of false statements, significantly improving the efficiency of personnel review and saving enterprises substantial costs. Furthermore, this patent has strong extensibility in the field of semantic parsing and discrimination, with a wide range of application scenarios and good feasibility.

[0148] This embodiment proposes a forum public opinion monitoring method that integrates a bidirectional hierarchical semantic interactive neural network and a deep learning model, aiming to efficiently identify abnormal public opinion and harmful information. The method first extracts and constructs a multi-branch chain-like propagation tree structure from the database to comprehensively present the propagation relationships between comments. In the feature extraction stage, TextCNN is used to extract semantic features from the content of each node, while convolutional operations between nodes capture the associated information in the propagation path, thereby obtaining local and contextual semantic features. To gain a deeper understanding of the propagation structure, this embodiment introduces GRU to perform sequential modeling of the multi-branch propagation path, learning the local features and global structural information of each branch. Furthermore, an attention mechanism is combined to dynamically weight the importance of different branches to highlight the influence of key propagation paths. Finally, the semantic information in the breadth direction (inter-node associations) and the depth direction (branch sequence) is fused and input into a fully connected layer for public opinion credibility classification. This embodiment, through comprehensive coverage of multi-dimensional semantic information, accurately characterizes the semantic features of information propagation, significantly improving the accuracy and reliability of public opinion monitoring, effectively enhancing the ability to distinguish between abnormal and normal public opinion, and providing strong support for high-precision public opinion analysis.

[0149] Example 2:

[0150] like Figure 6 As shown, this embodiment provides a forum public opinion monitoring device based on a bidirectional hierarchical semantic interaction neural network. The device includes:

[0151] Acquisition Unit 10 is used to acquire forum event data and construct a target propagation tree;

[0152] The input unit 20, connected to the acquisition unit 10, is used to input the target propagation tree into the bidirectional hierarchical semantic interaction neural network model to obtain the credibility classification result of the forum event, thereby realizing forum public opinion monitoring.

[0153] Among them, the bidirectional hierarchical semantic interaction neural network model is a deep learning model that outputs the credibility classification of forum public opinion events based on a tree-like propagation structure. It uses an interaction mechanism of top-down multi-branch semantic encoding and bottom-up global semantic representation, combined with the extraction of horizontal semantic features between peer nodes.

[0154] In one specific implementation, the device also includes a construction unit;

[0155] The building unit is connected to the input unit to construct a bidirectional hierarchical semantic interaction neural network model;

[0156] The building blocks specifically include:

[0157] The preprocessing module is used to collect historical tree-like propagation events and perform structured tree-like processing on these events to obtain tree-like propagation events; and to perform text preprocessing and abnormal public opinion labeling on the tree-like propagation events to form a tree-like sample set.

[0158] The partitioning module, connected to the preprocessing unit, is used to divide the tree-like sample set into training set and test set according to a preset ratio.

[0159] The training module, connected to the partitioning module, is used to iteratively train the initial semantic interaction neural network model based on the training set to obtain the initial classification and recognition model.

[0160] The testing module, connected to the training module, is used to input the test set into the initial classification and recognition model and obtain the test classification result for each test sample in the test set.

[0161] The adjustment module, connected to the testing module, is used to adjust the parameters of the initial semantic interaction neural network model based on the test classification results, thereby obtaining a bidirectional hierarchical semantic interaction neural network model.

[0162] The apparatus in this embodiment is capable of performing the method in Embodiment 1.

[0163] Example 3:

[0164] like Figure 7 As shown, this embodiment provides an electronic device, which includes a memory 200 and a processor 100. The memory 200 stores a computer program. When the processor 100 runs the computer program stored in the memory 200, the processor 100 executes the forum public opinion monitoring method based on a bidirectional hierarchical semantic interactive neural network as described in Embodiment 1.

[0165] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A forum public opinion monitoring method based on a bidirectional hierarchical semantic interaction neural network, characterized in that, The method includes the following steps: Acquire forum event data and construct a target propagation tree; The target propagation tree is input into a bidirectional hierarchical semantic interaction neural network model to obtain the credibility classification results of forum events, thereby realizing forum public opinion monitoring. The bidirectional hierarchical semantic interaction neural network model is a deep learning model that outputs the credibility classification of forum public opinion events based on a tree-like propagation structure. It uses an interaction mechanism of top-down multi-branch semantic encoding and bottom-up global semantic representation, combined with the extraction of horizontal semantic features between peer nodes.

2. The forum public opinion monitoring method based on a bidirectional hierarchical semantic interaction neural network according to claim 1, characterized in that, Before inputting the target propagation tree into the bidirectional hierarchical semantic interaction neural network model, the method further includes: constructing a bidirectional hierarchical semantic interaction neural network model; The construction of the bidirectional hierarchical semantic interaction neural network model specifically includes: Step A1: Collect historical tree-like propagation events and perform structured tree processing on the historical tree-like propagation events to obtain tree-like propagation events; and perform text preprocessing and abnormal public opinion labeling on the tree-like propagation events to form a tree-like sample set; Step A2: Divide the tree-structured sample set into a training set and a test set according to a preset ratio; Step A3: Based on the training set, iteratively train the initial semantic interaction neural network model to obtain the initial classification and recognition model; Step A4: Input the test set into the initial classification and recognition model to obtain the test classification result for each test sample in the test set; Step A5: Based on the test classification results, adjust the parameters of the initial semantic interaction neural network model to obtain a bidirectional hierarchical semantic interaction neural network model.

3. The forum public opinion monitoring method based on a bidirectional hierarchical semantic interaction neural network according to claim 2, characterized in that, Step A3 specifically includes: Step B1: Set the initial parameters of the initial semantic interaction neural network model; the initial parameters include batch size, stopping condition, and maximum number of iterations; Step B2: Input the training set into the initial semantic interaction neural network model in batches according to the batch size; Step B3: During the operation of the initial semantic interaction neural network model, an independent classification result is generated for each training sample in the training set based on the tree-like features; Step B4: Based on the classification results, update the model parameters of the initial semantic interaction neural network model and output the updated model parameters; Step B5: Repeat steps B2 to B4 until the maximum number of iterations is reached or the stopping condition is met; Step B6: Based on the updated model parameters output at the end, determine the initial semantic interaction neural network model as the initial interaction recognition model.

4. The forum public opinion monitoring method based on a bidirectional hierarchical semantic interaction neural network according to claim 2 or 3, characterized in that, Before step A3, the method further includes the step of: constructing an initial semantic interaction neural network model; The construction of the initial semantic interaction neural network model specifically includes: Construct a propagation tree; wherein the parent node of the propagation tree is the original post, and the child nodes of the propagation tree are comments; Both the parent node and the child node are encoded into sentence vectors; Based on the sentence vector, breadth information is extracted from comments at the same level to obtain a breadth information vector; and based on the sentence vector, semantics are transmitted from top to bottom and the parent node influences the child node to obtain a top-down information feature vector; and based on the sentence vector, information is aggregated from bottom to top, and all the child node information is summarized to the corresponding parent node to obtain a bottom-up feature vector. The breadth information vector, the top-down information feature vector, and the bottom-up feature vector are concatenated to obtain a multi-feature composite vector. The multi-feature integrated vector is input into a pre-selected classification learning model to obtain an initial semantic interaction neural network model.

5. The forum public opinion monitoring method based on a bidirectional hierarchical semantic interaction neural network according to claim 4, characterized in that, The breadth information vector is extracted using multiple convolutional kernels.

6. The forum public opinion monitoring method based on a bidirectional hierarchical semantic interaction neural network according to claim 4, characterized in that, The step of obtaining a top-down information feature vector based on the sentence vector, using semantic transmission from top to bottom and influencing the child nodes through the parent nodes, specifically includes: Determine the top-down transmission path; Based on the top-down transmission path, the sentence vector of the parent node and the sentence vector of the corresponding child node are semantically fused to obtain the first semantic information; If a parent node corresponds to multiple child node branches, then an attention mechanism is used to adjust and filter the semantic information of the sentence vector of the parent node and the sentence vector of the corresponding multi-branch child nodes. Summarize the first semantic information to obtain the top-down information feature vector; The step of aggregating information from bottom to top based on the sentence vector, summarizing all the child node information to the corresponding parent node, and obtaining a bottom-up feature vector, specifically includes: Determine the bottom-up transmission path; Based on the bottom-up transmission path, the sentence vectors of all child nodes under the same parent node are summarized and refined level by level to obtain the comprehensive semantic vector of the child nodes. The sentence vector of the parent node is semantically fused with the comprehensive semantic vector to obtain the second semantic information; By summarizing the second semantic information, a bottom-up feature vector is obtained.

7. The forum public opinion monitoring method based on a bidirectional hierarchical semantic interaction neural network according to claim 4, characterized in that, The step of concatenating the breadth information vector, the top-down information feature vector, and the bottom-up feature vector to obtain a multi-feature composite vector specifically includes: The top-down information feature vector and the bottom-up feature vector are concatenated to obtain the depth information vector; The breadth information vector and the depth information vector are concatenated to obtain a multi-feature composite vector.

8. A forum public opinion monitoring device based on a bidirectional hierarchical semantic interaction neural network, characterized in that, The device includes: The acquisition unit is used to acquire forum event data and construct a target propagation tree; The input unit, connected to the acquisition unit, is used to input the target propagation tree into the bidirectional hierarchical semantic interaction neural network model to obtain the credibility classification result of the forum event, thereby realizing forum public opinion monitoring. The bidirectional hierarchical semantic interaction neural network model is a deep learning model that outputs the credibility classification of forum public opinion events based on a tree-like propagation structure. It uses an interaction mechanism of top-down multi-branch semantic encoding and bottom-up global semantic representation, combined with the extraction of horizontal semantic features between peer nodes.

9. The forum public opinion monitoring device based on a bidirectional hierarchical semantic interaction neural network according to claim 8, characterized in that, The device also includes a construction unit; The building unit is connected to the input unit and is used to build a bidirectional hierarchical semantic interaction neural network model; The building unit specifically includes: The preprocessing module is used to collect historical tree-like propagation events and perform structured tree-like processing on the historical tree-like propagation events to obtain tree-like propagation events; and to perform text preprocessing and abnormal public opinion labeling on the tree-like propagation events to form a tree-like sample set. A partitioning module, connected to the preprocessing unit, is used to divide the tree-shaped sample set into a training set and a test set according to a preset ratio; The training module, connected to the partitioning module, is used to iteratively train the initial semantic interaction neural network model based on the training set to obtain an initial classification and recognition model. The testing module, connected to the training module, is used to input the test set into the initial classification and recognition model to obtain the test classification result for each test sample in the test set; An adjustment module, connected to the test module, is used to adjust the parameters of the initial semantic interaction neural network model based on the test classification results, thereby obtaining a bidirectional hierarchical semantic interaction neural network model.

10. An electronic device, characterized in that, The system includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the forum sentiment monitoring method based on a bidirectional hierarchical semantic interaction neural network according to any one of claims 1 to 7.

Images