Key group identification method for hot events

Abstract

The invention provides a key group identification method for a hot event. Comprising a clue discovery module triggered by a multi-source hotspot event and a group discovery module based on an event keyword graph; the input of the clue discovery module triggered by the multi-source hotspot event is multi-source text information on the social communication platform; detecting a hot event from the multi-source data stream through an element extraction model; then, the association between the keyword graph and the word graph is constructed by utilizing the event elements; when information increment occurs, the keyword graph is updated and divided, and the process of searching user information related to keywords is carried out; the implementation process of the group discovery module based on the event keyword graph comprises the following steps: constructing a social network graph; obtaining node representations of the user in all different social networks; using the user node representation and the sub-graph association region obtained in the step 2 as input, and inputting data into a micropooling network for processing multi-granularity information; and judging the sub-graphs so as to find an unknown group, namely a potential key group of the hot event.

Description

Technical Field

[0001] The present disclosure relates to the field of Internet, and more specifically, to a method for identifying key groups for hot events. Background Art

[0002] At present, with the advent of the Internet era, people are increasingly using social communication platforms to communicate in their daily lives. Many users form groups on social communication platforms to carry out messaging, file sharing and video conferencing functions, enabling individuals and teams to communicate and collaborate quickly and efficiently, reducing waiting time and speeding up information exchange, thereby improving work efficiency and productivity. At present, there are more and more hot entertainment events on social platforms, which not only enrich the content of social platforms, but also drive extensive discussion and participation of users. In order to better grasp the dissemination trend and user interaction pattern of entertainment hot spots, it is particularly important to actively identify groups related to these hot events. This not only helps the platform to better understand user interests and behaviors, but also provides data support and reference for promoting the positive development of entertainment hot spots. Current technical capabilities are mainly focused on the passive discovery of group and account information content. Since relevant groups often show high adaptability and variability, it is necessary to establish a flexible and adaptive key group discovery system to ensure accurate identification and discovery of target groups. It can be seen from this that a technical solution for key group identification for hot events is needed. Summary of the invention

[0003] The purpose of the embodiments of the present disclosure is to provide a method for identifying key groups for hot events, to discover key clues of groups from hot events, and to actively identify and discover key groups based on the group clues.

[0004] In a general aspect, a method for identifying key groups for hot events is provided, comprising a clue discovery module triggered by multi-source hot events and a group discovery module based on an event keyword graph;

[0005] The implementation process of the clue discovery module triggered by multi-source hot events is as follows: the input is multi-source text information on the social communication platform; the hot events are detected from the multi-source data stream through the element extraction model, and the selected event elements are defined as seven-tuples:

[0006] Event=<t,desc,locs,pars,key,emo,posts>

[0007] Among them, t is the timestamp of the event, indicating the time slice when the event occurred; desc is the description of the event, briefly summarizing the content and field of the event; locs is the set of related locations of the event, and generally there is only one location in the set to indicate where the event occurred; pars is the set of people or organizations involved in the event; key is the keyword describing the event; emo is the ideological tendency of the event; posts is the relevant reports of the event;

[0008] Then, event elements are used to construct keyword graphs and the associations between keyword graphs;

[0009] Afterwards, when information increment occurs, the keyword graph is updated and divided, and the process of searching for user information related to the keywords is performed;

[0010] The implementation process of the group discovery module based on the event keyword graph is as follows: Step 1, through the process of searching for user information involving keywords, searching on the social communication platform to obtain user-related data, and filtering out users related to the event, interaction relationships between users, and content posted by users through text clustering, element extraction, etc., to construct a social network graph;

[0011] Step 2: Obtain the node representation of the user in all different social networks, and use the optimization method based on subgraph modularity to extract the user's neighbor subgraph, that is, determine the subgraph association area related to a specific event, that is, screen the user groups closely related to the target event, the interaction behaviors between users, and the related content they publish. The specific method is: First, based on the social network graph constructed in step 1, the user's position in different social networks (i.e., nodes) and the degree of association with the target event, construct an initial subgraph. The nodes in the initial subgraph can be determined by the frequency of interaction between nodes. Then, the subgraph modularity P is calculated,

[0012]

[0013] Among them, A ij An adjacency matrix indicating whether there is an edge between nodes i and j (1 if there is an edge, 0 otherwise). i and k j is the degree of nodes i and j, i.e. the number of edges between them. m is the total number of edges in the graph. i ,c j) is an indicator function. If nodes i and j belong to the same subgraph (or community), its value is 1, otherwise it is 0. Next, a greedy algorithm is used to adjust the node allocation in the subgraph to gradually increase the modularity value, thereby forming a more cohesive subgraph. Through multiple iterations, it gradually converges to a relatively stable structure. During the modularity optimization process, highly associated nodes and edges are continuously maintained, and the most closely connected user groups and interactive behaviors between users are screened out as subgraphs related to specific events. And through the fusion method of integrating social behavior and content features, according to the user's social behavior, such as posting, forwarding, comment frequency, social interaction mode (such as friends with frequent interactions), etc., the user's behavior feature vector is obtained, and based on the specific keywords, emotions, themes and other features extracted from the content posted by the user, the content feature vector is generated by text processing technology (such as TF-IDF model in natural language processing). The behavior and content feature vectors are combined, and the final feature representation of the user is generated using a fusion method (such as weighted average or multi-layer neural network), ensuring that the user's behavior and content characteristics are retained at the same time, that is, the node representation of each user under a specific event. Then, step 1 is repeated multiple times, and each time the social network involved in different events is adjusted. The constructed Attention mechanism is used to weight the node representation of users under different events and their corresponding associated areas to obtain a global user node representation:

[0014]

[0015] where w i It is the weight for each event learned through the constructed Attention mechanism;

[0016] Step 3: Using the user node representation and subgraph associated area obtained in step 2 as input, the data is input into a differentiable pooling network for processing multi-granularity information. The differentiable pooling network for processing multi-granularity information specifically mines and learns the structural information in the existing sparse subgraph area by performing regional pooling operations on sparse subgraphs of different parts, wherein the subgraph embedding of the multi-granularity differentiable pooling is represented as:

[0017]

[0018] Among them, S diffPool represents a trainable pooling matrix, which is used to learn the uniform dimension of node feature vectors, A diffPool The adjacency matrix representing the coarse-grained pooling graph guides the screening and effective representation of multi-region joint pooling features;

[0019] Step 4: Send the differentiable pooled subgraph embedding into the binary classification discriminator to identify the key groups, discriminate the subgraphs, and thus discover unknown groups, which are the potential key groups of hot events.

[0020] The association between the word graphs is constructed using an incrementally trained Word2vec model, and it is necessary to connect nodes with strong word similarity or strong location association to construct a word graph semantic information subgraph. The specific steps are: first, use the initial word graph data to train a basic Word2Vec model. The model generates an initial vector representation of each word through a conventional Word2Vec training algorithm (such as Skip-gram). Then, in the incremental update phase, after receiving new word graph data or detecting new hot words in the semantic space, incremental training gradually adds new data to the model without changing the structure of the original word vector, specifically: 1) When new data is input, check whether there are new words in the vocabulary, and if so, add these words to the model. 2) Window update and context capture: Based on the sliding window, the context vocabulary around the new word is captured to adapt to the context distribution of the new word, and the vector representation of the related word is updated. Then, the edge condition is determined by the similarity of the word vectors (such as cosine similarity). If the semantic similarity between the two words is higher than a specific threshold, the edge is connected in the word graph. If there are words related to specific geographic locations or hot events in the new data (such as adjacent locations or places where events occurred), the model will learn the semantic relationship between these words and connect them in the word graph. Finally, an overlapping community discovery algorithm is performed on the semantic information subgraph of the word graph, that is, all k-cliques in the graph are detected, and cliques that share k-1 nodes are defined as adjacent. The largest set of them is aggregated together as a community. Each community is an event representative. The corresponding source data of the event representative is then restored from the index, and information is extracted for event representation. After removing duplications, a set of events within a time slice is obtained, and the keyword graph of a single source is fused into a keyword graph of multiple sources to achieve cross-source word graph alignment and update.

[0021] The constructed Attention mechanism is specifically a BERT-BiLSTM-CRF-IDCNN model, which is divided into three parts, namely, a BERT layer, a BiLSTM and ICDNN fusion layer, and a CRF layer;

[0022] First, the BERT pre-training layer is trained on a large-scale unlabeled dataset through an unsupervised method to extract rich grammatical and semantic features and obtain word vector representations; then, the trained word vectors are input into BiLSTM and IDCNN for feature extraction, and the output features of the two neural networks are merged; finally, the output features are input into the CRF layer for correction.

[0023] The BERT pre-training layer adopts the Transformer encoding structure, and the calculation method is:

[0024]

[0025] Where Q represents the query matrix, which indicates the degree of attention of the current word to other words. A query vector is generated for each word position to find words related to it. K represents the key matrix, which represents the characteristics or content of the word. A key vector is generated for each word position to match with the query vector (Q) to determine the relevance. V represents the value matrix, which represents the value or output information of the word. A value vector is generated for each word position, which represents the content information of the word, and will be weighted and summarized into the output. d k Represents the dimension of the key vector K, which is a normalization factor used to narrow the value range of the dot product result to avoid excessive values ​​causing the softmax function to approach saturation and lose discrimination. T represents the transposed matrix of the key matrix K. Multiple different linear mappings are performed on Q, K, and V, and then the new Q, K, and V are recalculated to obtain different attentions (Q, K, V), and then they are spliced ​​together. The calculation method using the "MultiHead" mechanism in the Transformer encoding structure is as follows:

[0026] MultiHead(Q,K,V)=Concat(Head1,Head2,...,Head n )W

[0027] Head i =Attention(QW i Q , KW i k , VW i V )

[0028] Among them, MultiHead(Q, K, V) represents the output of the multi-head attention mechanism. It concatenates the outputs of multiple attention heads to obtain a comprehensive representation. Concat means concatenating the outputs of all attention heads into a matrix by column. i Represents the output result of the i-th attention head, and each attention head calculates a separate attention. W represents the linear transformation weight matrix of the concatenated output matrix, which is used to map the concatenated result. Attention is the calculation formula of the self-attention mechanism, which represents the attention relationship between the current query vector, key vector and value vector. W i k , W i VRepresents the weight matrix corresponding to the i-th attention head, which is used to perform linear transformations on queries, keys, and values. Different attention heads use different weight matrices to learn different features. Residual networks and layer normalization are also added to Transformer to improve network degradation. The calculation method is as follows:

[0029]

[0030] FNN=max(0,xW1+b1)W2+b2

[0031] Among them, x i represents the i-th input feature, μ L Represents the mean of the current layer input x, which is used to eliminate mean shift, σ 2 Indicates the variance of the current layer input x, which is used to standardize the distribution of data. ε is a small constant to prevent the denominator from being zero. α and β are trainable scaling and offset parameters that help the model restore the expressiveness of the input features. In FNN, x is the input vector, W1 and W2 are the weight matrices of the feedforward network, and b1 and b2 are the bias vectors of the feedforward network. The max(0, ) function is the ReLU activation function, which is used to introduce nonlinearity to make the network more expressive.

[0032] In the BiLSTM module, LSTM introduces a gating mechanism, namely, input gate, output gate and forget gate, and uses memory cells to capture long-range dependencies between words. BiLSTM consists of two LSTMs in opposite directions to extract global context features.

[0033] The states of the three doors are updated as follows:

[0034] i t =σ g (W i x t +U i h t-1 +b i )

[0035] O t =σ g (W o x t +U o h t-1 +b i )

[0036]

[0037] where x t is the input of the current time step t, h t-1 is the hidden state of the previous step t-1, h t-1It is the forget gate, t is the input gate, O t is the output gate, b is the bias, is the information that the neuron remembers at time t, C t is the information that the current neuron wants to store, h t is the output of the final RNN unit, U i , U o , U c There are three weight matrices for input x. t Linear transformation of U. i , U o , U c There are three weight matrices, which are used to adjust the hidden state h of the previous time step. t-1 Perform a linear transformation, f t Represents the forget gate, which controls the degree of retention or forgetting of information in the memory unit. Represents element-by-element multiplication (Hadamard product), which is used to perform bit-wise calculations and merge the outputs of different gates. σg represents the sigmoid activation function, which maps the output value to [0, 1] and is used to control the switch state of each gate. h Represents the tanh activation function, which is used to generate the value of the candidate memory unit and map the output value to [-1, 1].

[0038] In the IDCNN module, the expansion width is added to the traditional CNN convolution, the data in the middle of the expansion width is skipped, and the size of the convolution kernel is kept unchanged for convolution operation.

[0039] In the CRF module, for a given sequence x = (x1, x2, x3, ...X n ) and the corresponding label sequence, after passing through the feature extraction layer of the neural network, y = (y1, y2, y3, ...y n ), the score of each label can be obtained by linear mapping:

[0040] P i =W s h t +b s

[0041] where h t is the input data x of the previous layer t Output at time t; Ws and Bs are linear mapping parameters; on this basis, CRF defines the label transfer score, and the score from the input sequence to the label sequence can be expressed as:

[0042]

[0043] Where W is the transformation matrix. i，jrepresents the label transfer score; P i，yi Represents the score of the character's yi label, W = (Wu, j) The maximum likelihood function of the training set {xi, yi} is:

[0044]

[0045] Where λ and θ are regularization parameters; P represents the probability corresponding to the original sequence to the predicted sequence:

[0046]

[0047] The innovation of the embodiment of the present invention lies in:

[0048] The present invention discovers key clues of groups from hot events, actively identifies and discovers key groups based on group clues, and the model design focuses on the following two aspects: a clue discovery module triggered by multi-source hot events and a group discovery module based on an event keyword graph. Extract hot event elements from the social communication platform, build a keyword graph, update group clues, and then discover specific users based on hot keywords. Through text clustering, element extraction, etc., screen out users related to the event, interactive relationships between users, and the content of user posts, build a social network graph, and then obtain the node representation of the user in all different social networks. Since the size of the group changes from small to large, with the evolution of multiple events, the interactive behaviors between and within groups gradually increase, making it easier to discover key groups.

[0049] 1. Facing the information processing needs of multiple platforms, existing group discovery research often only focuses on the group characteristics of a single source, while ignoring the impact of different source characteristics on identifying key groups. This invention integrates data of multiple dimensions by utilizing the implicit semantic information of hot events contained in multi-source data, which not only crosses the boundaries of platforms, but also provides more comprehensive and in-depth event analysis, thereby effectively improving the credibility and utilization of information.

[0050] 2. The present invention can quickly respond to hot events and update group clues in a timely manner, thereby greatly improving the timeliness and accuracy of event detection. This is particularly important because hot events are often accompanied by an explosive growth of a large amount of information, which contains a lot of noise. The present invention can effectively filter out valuable information from it.

[0051] 3. The present invention solves the problem of semantic drift caused by the passage of time by combining the incrementally trained Word2vec model, and strengthens the connection between nodes with similar meanings or location associations by fusing semantic-level subgraphs. This not only improves the accuracy of event discovery, but also can more accurately identify and fuse related events, avoiding the problem of information fragmentation caused by semantic dispersion. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The above and other objects and features of the present disclosure will become more apparent from the following description in conjunction with the accompanying drawings.

[0053] Figure 1 is a schematic diagram showing an architecture diagram of a method for identifying key groups for hot events according to an embodiment of the present disclosure;

[0054] Figure 2 is a schematic diagram showing a keyword graph evolution analysis based on event elements according to an embodiment of the present disclosure;

[0055] Figure 3 is a schematic diagram showing a BERT-BiLSTM-CRF-IDCNN architecture diagram according to an embodiment of the present disclosure;

[0056] Figure 4 Schematic diagram of a social network diagram according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0057] The following specific embodiments are provided to help the reader obtain a comprehensive understanding of the methods, devices and / or systems described herein. However, after understanding the disclosure of the present application, various changes, modifications and equivalents of the methods, devices and / or systems described herein will be clear. For example, the order of operations described herein is only an example and is not limited to those orders set forth herein, but can be changed as will be clear after understanding the disclosure of the present application, except for operations that must occur in a specific order. In addition, for greater clarity and simplicity, the description of features known in the art may be omitted.

[0058] The features described herein can be implemented in different forms and should not be construed as being limited to the examples described herein. Rather, the examples described herein have been provided to illustrate only some of the many possible ways to implement the methods, devices, and / or systems described herein, which will be clear after understanding the disclosure of the present application.

[0059] As used herein, the term "and / or" includes any one of the associated listed items and any combination of any two or more.

[0060] Although terms such as "first", "second", and "third" may be used herein to describe various members, components, regions, layers, or portions, these members, components, regions, layers, or portions should not be limited by these terms. Instead, these terms are only used to distinguish one member, component, region, layer, or portion from another member, component, region, layer, or portion. Therefore, without departing from the teachings of the examples described herein, the first member, first component, first region, first layer, or first portion referred to in the examples may also be referred to as the second member, second component, second region, second layer, or second portion.

[0061] In the specification, when an element (such as a layer, a region or a substrate) is described as being “on”, “connected to” or “coupled to” another element, the element may be directly “on”, “connected to” or “coupled to” another element, or one or more other elements may be present therebetween. Conversely, when an element is described as being “directly on”, “directly connected to” or “directly coupled to” another element, there may be no other elements present therebetween.

[0062] The terms used herein are only used to describe various examples and are not intended to limit the disclosure. Unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. The terms "comprise", "include" and "have" indicate the presence of the described features, quantities, operations, components, elements and / or combinations thereof, but do not exclude the presence or addition of one or more other features, quantities, operations, components, elements and / or combinations thereof.

[0063] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by a person of ordinary skill in the art to which the present disclosure belongs after understanding the present disclosure. Unless explicitly defined as such herein, terms (such as those defined in a general dictionary) should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and the present disclosure, and should not be interpreted in an idealized or overly formal manner.

[0064] Furthermore, in the description of examples, when it is considered that a detailed description of a well-known related structure or function would cause vague interpretation of the present disclosure, such a detailed description will be omitted.

[0065] Figure 1 It is a schematic diagram showing a method for identifying key groups for hot events according to an embodiment of the present disclosure.

[0066] The method is implemented by constructing two modules:

[0067] Clue discovery module triggered by multi-source hot events

[0068] The clue discovery module triggered by the multi-source hot events is input as multi-source text information on the social communication platform. Through the element extraction model, hot events are detected from the multi-source data stream, including the extraction of event elements such as time, description, related locations, participants, keywords, organizations, tendencies, etc., and then the event elements are used to construct a keyword graph, incrementally update and divide the keyword graph, analyze the evolutionary correlation of events, and continuously update group clues.

[0069] Since a large number of related discussions and news reports will erupt on social platforms in a short period of time after a hot event occurs, the massive amount of data contains a lot of noise, making it difficult to obtain comprehensive and highly reliable information from it. Existing research mainly focuses on event detection in a single channel, a single language or for a specific field, which loses some timeliness and comprehensiveness. For example, major events occurring abroad will be transmitted to the country with a certain delay, and multiple channels such as news media reports and social media channels are complementary in terms of information dissemination. To address this problem, the present invention uses implicit semantic information in multi-source data to connect data from multiple channels, and detect events from multi-source data streams in a timely and efficient manner. The present invention defines events as seven-tuples:

[0070] Event=<t,desc,locs,pars,key,emo,posts>

[0071] Here, t is the timestamp of the event, indicating the time slice when the event occurred; desc is the description of the event, briefly summarizing the content and field of the event; locs is the set of related locations of the event, and generally there is only one location in the set to indicate where the event occurred; pars is the set of people or organizations involved in the event; key is the keyword describing the event; emo is the ideological tendency of the event; posts is the related reports of the event, and the reports of each media type will be limited to a certain number, and only the most popular ones will be selected. Based on the seven-tuple definition of the event, the present invention detects hot events from multi-source data streams, including event elements such as time, description, related locations, participants, keywords, organizations, and tendencies.

[0072] Secondly, a keyword graph is constructed using multi-source text data streams. Specific weights are assigned to the nodes and edges in the graph based on the co-occurrence relationship between keywords. The changes in the weights of each node in the keyword graph are monitored to obtain the keyword graph. Figure 2The figure shows the general process of extracting word graphs from text streams and detecting events in a time sliding window. The left side of the figure shows a multi-source data stream, and the right side shows a constructed word graph. The number of word nodes in the actual event detection process is very large. The dark blue nodes in the figure refer to the detected key nodes, and the light blue nodes refer to the detailed description words that co-occur with the key nodes more times. The blue solid line indicates the co-occurrence between two word nodes, and its thickness indicates the frequency of word co-occurrence. The black dotted rectangle box indicates that the semantic distance between the two word nodes is very close. The blue dotted line indicates the edge connected by the fusion of semantic-level subgraphs. Two events are detected from this word graph, namely the four dark nodes above and the four dark nodes below.

[0073] After obtaining the keyword graph, in order to establish the association between the word graphs of the same event under different information, the present invention uses the incrementally trained Word2vec model to connect the nodes with strong similar meanings or strong location associations, and fuse the subgraphs based on the semantic space. The specific steps are: first, use the initial word graph data to train the basic Word2Vec model. The model generates the initial vector representation of each word through the conventional Word2Vec training algorithm (such as Skip-gram). Then, in the incremental update stage, after receiving new word graph data or detecting new hot words in the semantic space, the incremental training gradually adds the new data to the model without changing the structure of the original word vector, specifically: 1) When new data is input, check whether there are new words in the vocabulary, and if so, add these words to the model. 2) Window update and context capture: Based on the sliding window, the context vocabulary around the new word is captured to adapt to the context distribution of the new word, and the vector representation of the related word is updated. Then, the edge condition is determined by the similarity of the word vector (such as cosine similarity). If the semantic similarity between the two words is higher than a specific threshold, the edge is connected in the word graph. If the associated words of a specific geographical location or hot event appear in the new data (such as adjacent locations or places where events occurred), the model will learn the semantic relationship of these words and connect the edges in the word graph. Then before running the overlapping community discovery algorithm, the present invention removes the edges with edge weights less than or equal to 1, and performs the overlapping community algorithm on the abnormal subgraph that is reconnected to obtain a set of events in a specific time period. Specifically, all k-cliques in the graph are detected, and cliques that share k-1 nodes are defined as adjacent. The largest set obtained by aggregating together is a community, and each community is an event representative. After that, the corresponding source data of the event representative is restored from the index, and time, description, location, participants, keywords, fields, sensitivity, heat and other information are extracted for event representation. After removing duplications, a set of events within a time slice is obtained, and the keyword graph of a single source is merged into a keyword graph of multiple sources to achieve word graph alignment and update across sources.

[0074] Group discovery module based on event keyword graph

[0075] The group discovery module based on the event keyword graph searches on the social communication platform to obtain data such as user posts, and selects users related to the event, the interaction relationship between users, and the content of user posts through text clustering, element extraction, etc., to build a social network graph, and then obtain the node representation of the user in all different social networks. Since the size of the group changes from small to large, with the evolution of multiple events, the interaction behaviors between and within groups gradually increase, so that the key groups can be judged more accurately.

[0076] In the initial stage of social network analysis, the first task is to obtain the node representation of users in specific social scenarios from the complex social network graph. The social network graph is a complex that comprehensively reflects all the information of the social network, including the interaction relationship between users, published content, and social connections. This step is based on the optimization method of subgraph modularity to extract the user's neighbor subgraph, that is, to determine the associated area related to a specific event. This involves screening user groups that are closely related to the target event, the interaction behaviors between these users, and the related content they publish. On this basis, through the fusion detection method that integrates social behavior and content features, the node representation of each user in a specific event is accurately obtained.

[0077] Subsequently, in order to obtain a user node representation from a global perspective, the first step needs to be repeated multiple times, but each time it is adjusted for the social networks involved in different events. This process allows us to capture the behavioral dynamics of the same user node under different events and reflect its behavioral characteristics in diverse social scenarios. Using the Attention mechanism, the node representation of users under different events and their corresponding associated regions can be effectively weighted to obtain a global user node representation:

[0078]

[0079] Here, w i The weight for each event is learned through the Attention mechanism, which ensures that the model can focus on more informative event features.

[0080] Finally, in the face of ever-changing and growing social groups, especially during the evolution of multiple events, the increase in interactions between and within groups makes it easier to discover key groups. Using the user node representation and subgraph association area obtained in the previous step as input, this step feeds the data into a differentiable pooling network (DIFFPOOL) for processing multi-granularity information. DIFFPOOL can effectively mine and learn the structural information in the existing sparse subgraph area by performing regional pooling operations on different parts of the sparse subgraph. The representation of the subgraph embedding of the differentiable pooling is:

[0081]

[0082] Among them, S diffPool represents a trainable pooling matrix, which is used to learn the uniform dimension of node feature vectors, A diffPool The adjacency matrix representing the coarse-grained pooling graph guides the screening and effective representation of multi-region joint pooling features. Finally, the differentiable pooled subgraph embedding is fed into the binary classification discriminator to identify key groups.

[0083] In a more specific embodiment, the hot search event "A public figure getting a certificate" on the social platform is selected to specifically explain the entire implementation process. Figure 1 As shown, the present invention provides a method and device for identifying key groups for hot events. In this example, the input is official news on a social platform, and the output is groups related to hot events. The method includes two modules:

[0084] Clue discovery module triggered by multi-source hot events: The input of this module is multi-source text information on the social communication platform. Through the feature extraction model, hot events are detected from the multi-source data stream, including the extraction of event elements such as time, description, related locations, participants, keywords, organizations, tendencies, etc., and then the event elements are used to build a keyword graph, incrementally update and divide the keyword graph, analyze the evolutionary correlation of events, and continuously update group clues.

[0085] Since a hot event will cause a large number of related discussions and news reports on social platforms in a short period of time, and the massive data contains a lot of noise, the present invention uses the named entity extraction technology based on the large model BERT to extract the event seven-tuple. This method uses the overall structure of the BERT-BiLSTM-CRF-IDCNN model as shown in the figure. Figure 3As shown. The whole model is divided into three parts, namely the BERT layer, the BiLSTM and ICDNN fusion layer, and the CRF layer. First, the BERT pre-training layer is trained on a large-scale unlabeled data set through an unsupervised method to extract rich grammatical and semantic features and obtain word vector representations. Then, the trained word vectors are input into BiLSTM and IDCNN for feature extraction, and the output features of the two neural networks are merged. Finally, the output features are input into the CRF layer for correction. Specifically, the present invention first enters the BERT model training. The calculation method is shown in the formula.

[0086]

[0087] The "MultiHead" mechanism is used in the Transformer encoding structure, which can capture information in multiple dimensions and improve the representation ability of word vectors. Multiple different linear mappings are performed on Q, K, and V, and then the new Q, K, and V are recalculated to obtain different attentions (Q, K, V), and then they are spliced ​​together. The calculation method of the "MultiHead" mechanism is shown in the formula:

[0088] MultiHead(Q,K,V)=Concat(Head1,Head2,...,Head n )W

[0089] Head i =Attention(QW i Q , KW i k , VW i V )

[0090] Residual networks and layer normalization are also added to Transformer to improve network degradation. The calculation method is shown in the following formula:

[0091]

[0092] FNN=max(0,xW1+b1)W2+b2

[0093] In the formula, α and β are the parameters that need to be learned. μ and σ are the mean and variance of the input layer. These parameters are used in the layer normalization process. By learning and adjusting them, the model can better adapt to different input data and improve the training effect of the network.

[0094] In the Bidirectional Long Short-Term Memory (BiLSTM) module, LSTM introduces a gating mechanism, namely the input gate, output gate, and forget gate, using memory cells to capture long-range dependencies between words. BiLSTM consists of two LSTMs in opposite directions to extract global context features. The state of the three gates is updated as follows:

[0095] i t =σ g (W i x t +U i h t-1 +b i )

[0096] O t =σ g (W o x t +U o h t-1 +b i )

[0097]

[0098] where x t is the current input, h t-1 is the hidden state of the previous step, h t-1 It is the forget gate, t is the input gate, O t is the output gate, b is the bias, is the information that the neuron remembers at time t, C t is the information that the current neuron wants to store. t is the output of the final RNN unit.

[0099] In the Iterative Dilated Convolutional Neural Network (IDCNN) module, GPU resources can be fully utilized while using a simple structure similar to LSTM to remember as much input information as possible. IDCNN adds a dilation width to the traditional CNN convolution, skips the data in the middle of the dilation width, and keeps the size of the convolution kernel unchanged for convolution operations.

[0100] In the CRF module, in order to avoid the situation where some labels appear continuously and do not conform to the language logic, CRF can consider the logical relationship between labels to obtain the global optimal label sequence, and use it in the last layer of the model to further correct the recognition results. The principle is as follows: for a given sequence x = (x1, x2, x3, ...X n ) and the corresponding label sequence, after passing through the feature extraction layer of the neural network, y = (y1, y2, y3, ...y n ), the score of each label can be obtained by linear mapping:

[0101] P i =W s h (t) +b s

[0102] Where h(t) is the output of the previous layer's input data x(t) at time t. Ws and Bs are linear mapping parameters. On this basis, CRF defines the label transfer score, and the score from the input sequence to the label sequence can be expressed as:

[0103]

[0104] Where W is the transformation matrix. Wi,j represents the label transfer score. Pi,yi represents the score of the character's yi label. The maximum likelihood function of W = (Wu,j) training set {xi,yi} is:

[0105]

[0106] Where λ and θ are regularization parameters. P represents the probability corresponding to the original sequence to the predicted sequence:

[0107]

[0108] Based on the above method, get the specific content of the event seven-tuple:

[0109] Event=<Time: Date and time when the event occurred.

[0110] Description: The details of the incident, including what happened and how it happened.

[0111] Relevant locations: XX, etc.

[0112] Participants: Public figure A, Hong, etc.

[0113] Keywords: "A public figure", "love brain, A public figure", "pregnancy, A public figure Hong got married, A public figure, A public figure fans' current situation belike" and so on.

[0114] Organization: XX Hospital, A public figure fan club.

[0115] Tendency: envy, blessing, etc. >

[0116] Similarly, other text content will also be extracted through the BERT model. In this way, the key element information related to the pre-defined seven-tuple in the sentence is finally obtained.

[0117] Figure 2The paper shows the process of extracting word graphs from text streams in a time sliding window, aligning and fusing them based on the keyword graph information in different time periods using the overlapping community discovery algorithm, and detecting the general process of events. First, an initial time element graph G = (V, E) is defined, where V represents the node set in the graph, i.e., the keywords extracted from the text, including (A public figure, marriage, etc.) in this example; E represents the edge set, i.e., based on the co-occurrence relationship between keywords. For each pair of nodes v in the graph, i and v j , if they co-occur in the text, there is an edge e between them ij ∈E. The edge weight represents the frequency of co-occurrence between keywords. The co-occurrence frequency f ij To represent the edge e ij The weight of ij )=f ij , where w(e ij ) represents edge e ij Then, the present invention realizes the keyword graph G by monitoring the weight changes of the nodes. k If the weight of each node is defined as the sum of the weights of all edges of the node, that is, w(v i )=∑ j w(e ij ), where w(v i ) is the node v i The weight of the node, then the abnormal change of the node weight may indicate the occurrence of an event. The semantic distance between word nodes can be further defined as d ij Represents node v i and v j If d ij If the value is less than a certain threshold, the two nodes are semantically very close and can be represented by a black dotted rectangle. In addition, the dark blue nodes in the figure refer to the detected abnormal word nodes, the light blue nodes refer to the detailed description words that co-occur with the abnormal word nodes more times, and the blue solid line indicates the co-occurrence of the two word nodes, and its thickness indicates the frequency of word co-occurrence. The blue dotted line represents the edge connected by the fusion of the semantic level subgraph.

[0118] After getting the keyword graph G k Finally, in order to establish the association between word graphs of the same event under different information, it is necessary to combine the incrementally trained Word2vec model to connect the nodes with strong similarity in meaning or strong association in location, and fuse the subgraphs based on the semantic space. i and v j The semantic similarity between them is sim(v i , v j ), where sim(v i , vj )∈[0,1]. If sim(v i , v j ) is greater than a predetermined threshold, such as 0.8, then the edge between the two nodes indicates that they are strongly related in semantics or location. That is, if [sim(v i , v j )>θ, then connect v i and v j After the semantic subgraphs are fused, the present invention removes the edges whose weight is less than or equal to 1 [for any edge e ij , if w(e ij )≤1, then remove e ij ], the overlapping community algorithm is performed on the reconnected abnormal subgraph to obtain the event set at time. Subsequently, a k-clique is defined as a subgraph containing k interconnected nodes. If two k-cliques share k-1 nodes, they are defined as adjacent. By aggregating adjacent k-cliques, a large community can be found, which is a representative of an event. It can be expressed mathematically as:

[0119] - Find all k-cliques.

[0120] - Define k-cliques that share k-1 nodes as neighbors, and aggregate neighboring k-cliques to get communities.

[0121] The corresponding source data of the event representative is then restored from the index, and information such as time, description, location, participants, keywords, fields, sensitivity, and popularity are extracted for event representation. After removing duplicates, a set of events within a time slice is obtained.

[0122] The group discovery module based on the event keyword graph searches on the social communication platform. In this example, we search on the social platform with keywords such as "wedding" and "celebrity" to obtain user postings (#A public figure is in love# If you really like B, shouldn't you bless him? Even if he is a scumbag, I have the capital to try and make mistakes [lightly] When you love, just love boldly. Why don't you get married at the age of 35? How much money should you make to be considered a lot? Really, if you change to an ordinary person, I will congratulate A public figure#). Through text clustering and feature extraction, we can filter out users related to the event, the interaction relationship between users, and the content of user posts, such as posting tags (#A public figure is really... too in love#). We search based on the tags and find a group of users who have made relevant remarks. We build a social network graph with each user ID as a node, such as Figure 4As shown in Figure 2. Then, the node representation of the user in all different social networks is obtained. Since the size of the group changes from small to large, with the evolution of multiple events, the interaction behaviors between and within groups gradually increase, so the size of the social network graph continues to increase.

[0123] In the initial stage of social network analysis, the first task is to obtain the node representation of users in specific social scenarios from the complex social network graph. The social network graph is a complex that comprehensively reflects all the information of the social network, including the interaction relationship between users, published content, and social connections. This step is based on the optimization method of subgraph modularity to extract the user's neighbor subgraph, that is, to determine the associated area related to a specific event. This involves screening user groups that are closely related to the target event, the interaction behaviors between these users, and the related content they publish. On this basis, through the fusion detection method that integrates social behavior and content features, the node representation of each user in a specific event is accurately obtained.

[0124] Subsequently, in order to obtain a user node representation from a global perspective, the first step needs to be repeated multiple times, but each time it is adjusted for the social networks involved in different events. This process allows us to capture the behavioral dynamics of the same user node under different events and reflect its behavioral characteristics in diverse social scenarios. Using the Attention mechanism, the node representation of users under different events and their corresponding associated regions can be effectively weighted to obtain a global user node representation:

[0125]

[0126] Here, w i The weight for each event is learned through the Attention mechanism, which ensures that the model can focus on more informative event features.

[0127] Finally, in the face of ever-changing and growing social groups, especially during the evolution of multiple events, the increase in interactions between and within groups makes unknown groups easier to discover. Using the user node representation and subgraph association area obtained in the previous step as input, this step feeds the data into a differentiable pooling network (DIFFPOOL) for processing multi-granularity information. DIFFPOOL can effectively mine and learn the structural information in the existing sparse subgraph area by performing regional pooling operations on different parts of the sparse subgraph. The representation of the subgraph embedding of the differentiable pooling is:

[0128]

[0129] Among them, S diffPoolrepresents a trainable pooling matrix, which is used to learn the uniform dimension of node feature vectors. diffPool The adjacency matrix representing the coarse-grained pooling graph guides the screening and effective representation of multi-region joint pooling features. Finally, the differentiable pooled subgraph embedding is fed into the binary classification discriminator to identify key groups.

[0130] Although some embodiments of the present disclosure have been shown and described, it will be appreciated by those skilled in the art that modifications may be made to the embodiments without departing from the principles and spirit of the present disclosure, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for identifying key groups for hot events, characterized in that: It includes a clue discovery module triggered by multi-source hot events and a group discovery module based on event keyword graphs; The implementation process of the clue discovery module triggered by multi-source hot events is as follows: the input is multi-source text information on the social communication platform; the hot events are detected from the multi-source data stream through the element extraction model, and the selected event elements are defined as seven-tuples: Event=<t,desc,locs,pars,key,emo,posts> Where t is the timestamp of the event, indicating the time slice when the event occurs; desc is the description of the event, briefly summarizing the content and scope of the event; locs is a collection of related locations of the event, usually there is only one location in the collection to represent the place where the event occurred; Pars is a collection of people or organizations involved in an event; key is a keyword describing an event; emo is the ideological tendency of an event; posts is related reports on an event; Then, event elements are used to construct keyword graphs and the associations between keyword graphs; Afterwards, when information increment occurs, the keyword graph is updated and divided, and the process of searching for user information related to the keywords is performed; The implementation process of the group discovery module based on the event keyword graph is as follows: Step 1, through the process of searching for user information involving keywords, searching on the social communication platform to obtain user-related data, and filtering out users related to the event, interaction relationships between users, and content posted by users through text clustering, element extraction, etc., to construct a social network graph; Step 2: Obtain the node representation of the user in all different social networks, and use the optimization method based on subgraph modularity to extract the user's neighbor subgraph, that is, determine the subgraph association area related to a specific event, and screen the user groups closely related to the target event, the interaction behaviors between users, and the related content they publish; the specific method is: first, based on the social network graph constructed in step 1, the user's position in different social networks and the degree of association with the target event, construct an initial subgraph; determine the nodes in the initial subgraph by the interaction frequency between nodes, and then calculate the subgraph modularity P, Among them, A ij The adjacency matrix indicating whether there is an edge between nodes i and j; k i and k j is the degree of nodes i and j, i.e. the number of edges between them. m is the total number of edges in the graph. i ,c j ) is an indicator function. If nodes i and j belong to the same subgraph, its value is 1, otherwise it is 0; then, a greedy algorithm is used to adjust the node allocation in the subgraph to gradually increase the modularity value, thereby forming a more cohesive subgraph. Through multiple iterations, it gradually converges to a more stable structure. During the modularity optimization process, highly associated nodes and edges are continuously maintained, and the closest user groups and interactive behaviors between users are screened out as subgraphs related to specific events; and through a fusion method that integrates social behavior and content features, based on the user's social behavior, the user's behavior feature vector is obtained, and specific keywords, emotions, and topic features are extracted from the content posted by the user, and a content feature vector is generated using text processing technology; the behavior and content feature vectors are combined, and the fusion method is used to generate the user's final feature representation, ensuring that the user's behavior and content characteristics are retained at the same time, that is, the node representation of each user under a specific event. Then, step 1 is repeated multiple times, and each time the social network involved in different events is adjusted. The constructed Attention mechanism is used to weight the node representation of users under different events and their corresponding associated areas to obtain a global user node representation: where w i It is the weight for each event learned through the constructed Attention mechanism; Step 3: Using the user node representation and subgraph associated area obtained in step 2 as input, the data is input into a differentiable pooling network for processing multi-granularity information. The differentiable pooling network for processing multi-granularity information specifically mines and learns the structural information in the existing sparse subgraph area by performing regional pooling operations on sparse subgraphs of different parts, wherein the subgraph embedding of the multi-granularity differentiable pooling is represented as: Among them, S diffPool represents a trainable pooling matrix, which is used to learn the uniform dimension of node feature vectors, A diffPool The adjacency matrix representing the coarse-grained pooling graph guides the screening and effective representation of multi-region joint pooling features; Step 4: Send the differentiable pooled subgraph embedding into the binary classification discriminator to identify the key groups, discriminate the subgraphs, and thus discover unknown groups, which are the potential key groups of hot events.

2. A method for identifying key groups for hot events as claimed in claim 1, characterized in that: The association between the word graphs is constructed using an incrementally trained Word2vec model, and it is necessary to connect nodes with strong similar meanings or strong location associations to construct a word graph semantic information subgraph. The specific steps are: First, use the initial word graph data to train a basic Word2Vec model. The model generates an initial vector representation for each word through a conventional Word2Vec training algorithm; then, in the incremental update phase, after receiving new word graph data or detecting new hot words in the semantic space, incremental training gradually adds new data to the model without changing the structure of the original word vector, specifically: 1) When new data is input, check whether there are new words in the vocabulary. If so, add these words to the model; 2) Window update and context capture: Capture the contextual vocabulary around the new word based on a sliding window to adapt to the contextual distribution of the new word, and update the vector representation of the related words; then, determine the edge conditions based on the similarity of the word vectors. If the semantic similarity between the two words is higher than a certain value, the edge condition is determined. If a threshold is set, edges are connected in the word graph. If related words of a specific geographic location or hot event appear in the new data, the model learns the semantic relationship between these words and connects them in the word graph. Finally, an overlapping community discovery algorithm is performed on the semantic information subgraph of the word graph, that is, all k-cliques in the graph are detected, and cliques that share k-1 nodes are defined as adjacent. The largest set of cliques aggregated together is a community. Each community represents an event. The corresponding source data of the event representative is then restored from the index, and information is extracted for event representation. After removing duplications, a set of events within a time slice is obtained, and the keyword graph of a single source is fused into a keyword graph of multiple sources to achieve word graph alignment and update across sources.

3. A method for identifying key groups for hot events as claimed in claim 2, characterized in that: The constructed Attention mechanism is specifically a BERT-BiLSTM-CRF-IDCNN model, which is divided into three parts, namely, a BERT layer, a BiLSTM and ICDNN fusion layer, and a CRF layer; First, the BERT pre-training layer is trained on a large-scale unlabeled dataset using an unsupervised method to extract rich grammatical and semantic features and obtain word vector representations. Then, the trained word vectors are input into BiLSTM and IDCNN for feature extraction, and the output features of the two neural networks are merged. Finally, the output features are input into the CRF layer for correction.

4. A method for identifying key groups for hot events as claimed in claim 3, characterized in that: The BERT pre-training layer adopts the Transformer encoding structure, and the calculation method is: Where Q represents the query matrix, which indicates the degree of attention of the current word to other words. A query vector is generated for each word position to find related words. K represents the key matrix, which indicates the characteristics or content of the word. A key vector is generated for each word position to match with the query vector Q to determine the relevance. V represents the value matrix, which indicates the value or output information of the word. A value vector is generated for each word position to indicate the content information of the word, which will be weighted and summarized into the output. k Represents the dimension of the key vector K, which is a normalization factor used to narrow the value range of the dot product result to avoid excessive values ​​causing the softmax function to approach saturation and thus lose discrimination. T represents the transposed matrix of the key matrix K. Multiple different linear mappings are performed on Q, K, and V, and then the new Q, K, and V are recalculated to obtain different attentions (Q, K, V), and then they are spliced ​​together. The calculation method using the "MultiHead" mechanism in the Transformer encoding structure is as follows: MultiHead(Q,K,V)=Concat(Head1,Head2,…,Head n )W Among them, MultiHead(Q, K, V) represents the output of the multi-head attention mechanism. After concatenating the outputs of multiple attention heads, a comprehensive representation is obtained. Concat means concatenating the outputs of all attention heads into a matrix by column. i Represents the output result of the i-th attention head. Each attention head calculates a separate attention. W represents the linear transformation weight matrix of the concatenated output matrix, which is used to map the concatenated result. Attention is the calculation formula of the self-attention mechanism, which represents the attention relationship between the current query vector, key vector and value vector. Represents the weight matrix corresponding to the i-th attention head, which is used to perform linear transformation on queries, keys, and values. Different attention heads use different weight matrices to learn different features. Residual networks and layer normalization are also added to the Transformer to improve network degradation. The calculation method is as follows: FNN=max(0,xW1+b1)W2+b2 Among them, x i represents the i-th input feature, μ L Represents the mean of the current layer input x, which is used to eliminate mean shift, σ 2 It represents the variance of the current layer input x, which is used to standardize the distribution of data. ε represents a small constant to prevent the denominator from being zero. α and β represent trainable scaling and offset parameters to help the model restore the expressiveness of the input features. In FNN, x is the input vector, W1 and W2 represent the weight matrix of the feedforward network, b1 and b2 represent the bias vector of the feedforward network, and the max(0,·) function is the ReLU activation function, which is used to introduce nonlinearity to make the network more expressive.

5. A method for identifying key groups for hot events as claimed in claim 4, characterized in that: In the BiLSTM module, LSTM introduces a gating mechanism, namely, input gate, output gate and forget gate, and uses memory cells to capture long-range dependencies between words. BiLSTM consists of two LSTMs in opposite directions to extract global context features. The states of the three doors are updated as follows: i t =σ g (W i x t +U i h t-1 +b i ) The t =s g (W o x t +U o h t-1 +b i ) where x t is the input of the current time step t, h t-1 is the hidden state of the previous step t-1, h t-1 It is the forget gate, t is the input gate, O t is the output gate, b is the bias, is the information that the neuron remembers at time t, C t is the information that the current neuron wants to store, h t is the output of the final RNN unit, U i , U o , U c There are three weight matrices for input x. t Linear transformation, U i , U o , U c There are three weight matrices, which are used to adjust the hidden state h of the previous time step. t-1 Perform a linear transformation, f t Represents the forget gate, which controls the degree of retention or forgetting of information in the memory unit. Represents element-by-element multiplication (Hadamard product), which is used to perform bitwise calculations and merge the outputs of different gates, σ g Represents the sigmoid activation function, which maps the output value to [0,1] and is used to control the switch state of each gate. h Represents the tanh activation function, which is used to generate the value of the candidate memory unit and map the output value to [-1,1].

6. A method for identifying key groups for hot events as claimed in claim 5, characterized in that: In the IDCNN module, the expansion width is added to the traditional CNN convolution, the data in the middle of the expansion width is skipped, and the size of the convolution kernel is kept unchanged for convolution operation.

7. A method for identifying key groups for hot events as claimed in claim 6, characterized in that: In the CRF module, for a given sequence x = (x1, x2, x3, ... x n ) and the corresponding label sequence, after passing through the feature extraction layer of the neural network, y=(y1,y2,y3,…y n ), the score of each label can be obtained by linear mapping: P i =W s h t +b s Its t is the input data x of the previous layer t Output at time t; W s and B s is a linear mapping parameter; on this basis, CRF defines the label transfer score, and the score from the input sequence to the label sequence can be expressed as: Where W is the transformation matrix, W i，j represents the label transfer score; P i，yi Represents the score of the character's yi label, W = (Wu, j) The maximum likelihood function of the training set {xi, yi} is: Where λ and θ are regularization parameters; P represents the probability corresponding to the original sequence to the predicted sequence:

Images