News event detection method and device based on multi-dimensional heterogeneous graph representation and matching

Abstract

The application discloses a news event detection method and device based on multi-dimensional heterogeneous graph representation and matching, solves the problems of unified modeling and embedding of various attributes of news data, the problems of difficult matching between graph structures and news similarity calculation, the problems of incapability of modeling dynamic changes of events over time and incapability of identifying related events in news event detection based on graphs, and the problem of low efficiency of news event detection based on graphs, and can accurately detect news events in different fields. The method comprises the following steps: (1) extracting main features of news data as node information, and constructing a news multi-dimensional heterogeneous graph; (2) learning embedded values of the news multi-dimensional heterogeneous graph, and iteratively training to generate optimal model parameters; (3) using a graph neural network to match the news multi-dimensional heterogeneous graph, and calculating a matching score between news; and (4) creating a dynamic graph model for news data, using a subgraph in a time interval covered by a sliding time window, and dynamically detecting news events.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and in particular to a news event detection method based on multi-dimensional heterogeneous graph representation and matching, and a news event detection device based on multi-dimensional heterogeneous graph representation and matching. BACKGROUND

[0002] With the rapid progress of the Internet and various news media platforms, people's access to information is increasingly diversified. People can access the latest news and information from all over the world at any time and any place through smart phones and tablets, whether it is international political dynamics, economic development, technological innovation breakthroughs or popular cultural trends. The massive news resources provided by these platforms greatly enrich and facilitate people's daily life. However, while bringing convenience to people's life, it also brings many challenges. The most prominent challenge is the increasingly serious information overload, and people often have difficulty quickly identifying valuable content in the vast information flood. This information overload not only consumes a lot of time and effort of users, but also may lead to misreading and misunderstanding of information, making it more difficult to distinguish between true and false and grasp the key points, which has become a difficult problem that users generally concern. News event detection is an effective solution to this problem, which uses advanced computer algorithms and big data analysis techniques to automatically identify, classify and track the occurrence and development of news events. Through real-time monitoring and intelligent analysis of a large amount of news data, this technology can help regulatory authorities, media organizations and even ordinary users quickly identify hot events, thereby effectively alleviating the pressure brought by information overload. News event detection technology not only improves the efficiency of information processing, but also helps to build a healthier and more orderly network information environment.

[0003] At present, the research in the field of news event detection includes methods based on traditional features, methods based on topic models, and methods based on graphs. Among them, the method based on graph is the most popular news event detection method at present. However, since the event detection method based on graph mostly relies heavily on linguistic knowledge and semantic features of news content, it brings problems such as information loss, semantic sparsity, and inaccurate event detection. And this kind of method needs to use community detection algorithm to divide the graph, but since the community detection algorithm has high time complexity, therefore under large-scale data, the event detection method based on graph has large time overhead, which limits its application range. SUMMARY

[0004] To overcome the defects of the prior art, the technical problem to be solved by the present application is to provide a news event detection method based on multi-dimensional heterogeneous graph representation and matching, which solves the problems of unified modeling and embedding of various attributes of news data, the difficulty in matching between graph structures and the similarity calculation of news, the inability to model the dynamic changes of events over time and the inability to identify related events in the news event detection based on graphs, and the low efficiency of the news event detection based on graphs, and can accurately detect news events in different fields and effectively improve the efficiency of news event detection.

[0005] The technical solution of the present application is that the news event detection method based on multi-dimensional heterogeneous graph representation and matching comprises the following steps:

[0006] (1) extracting the main features of news data, including theme features, time features and geographic location features, regarding them as node information, constructing a news multi-dimensional heterogeneous graph, and inputting part of the training data set;

[0007] (2) learning the embedding value of the established news multi-dimensional heterogeneous graph, and finally generating the optimal model parameters through step-by-step iterative training;

[0008] (3) using a graph neural network to match the constructed news multi-dimensional heterogeneous graph and calculate the matching score between news;

[0009] (4) dynamically detecting news events, first creating a dynamic graph model for news data, and then independently detecting events using the subgraph in the time interval covered by the sliding time window.

[0010]

[0011] ​This invention takes into account multiple pieces of news information, including topic, time, and location, fully utilizing key information within the news and establishing relationships between these key pieces of information. This method offers stronger news representation capabilities than traditional methods. The proposed method for matching news with multidimensional heterogeneous graphs uses heterogeneous graph embedding technology to obtain embedding values ​​for nodes in the heterogeneous graph. Based on this, a matching network is used to encode node pairs in the graph to obtain matching vectors. A graph convolutional network is then used to aggregate these local matching vectors into a final matching score for two news articles. This addresses the difficulty of matching graph models in non-Euclidean spaces and enables more accurate comparison of similarities between news articles. The proposed method for news event detection based on dynamic graph models uses a sliding time window Δt to partition a multidimensional heterogeneous news graph sequence and independently performs event detection on each heterogeneous graph sequence within each time window. This method enables more accurate detection of news events and identification of relationships between events, effectively improving the efficiency of news event detection. The proposed method can be used for news event detection and evolution, enabling regulatory authorities to gain a clearer understanding of the development patterns and evolutionary patterns of events, and thus more accurately predict the trends of event development. In today's society with massive data, monitoring key news and public opinion management are of great significance.

[0012] A news event detection device based on multi-dimensional heterogeneous graph representation and matching is also provided, which includes:

[0013] The news multi-dimensional heterogeneous graph modeling module is configured to extract the main features of news data, including topic features, time features, and geographic location features, and regard them as node information.

[0014] Information is collected and a multi-dimensional heterogeneous graph of news is constructed, with the input part being the training data set;

[0015] The news multi-dimensional heterogeneous graph embedding training module is configured to learn the embedding value of the established news multi-dimensional heterogeneous graph, and finally generate the optimal model parameters through step-by-step iterative training;

[0016] The news multi-dimensional heterogeneous graph representation matching module is configured to use a graph neural network to match the constructed news multi-dimensional heterogeneous graph and calculate the matching scores between news items;

[0017] The news event detection module based on the dynamic graph model is configured to dynamically detect news events. First, a dynamic graph model is created for the news data, and then subgraphs within the time interval covered by the sliding time window are used to independently detect events. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flowchart of a news event detection device based on multi-dimensional heterogeneous graph representation and matching according to the present invention.

[0019] Figure 2 A news data modeling model based on multi-dimensional heterogeneous graph according to the application is shown.

[0020] Figure 3 A news multi-dimensional heterogeneous graph embedding training flow chart according to the application is shown.

[0021] Figure 4 A news representation matching flow chart based on multi-dimensional heterogeneous graph according to the application is shown.

[0022] Figure 5 A news event detection flow chart based on a dynamic graph model according to the application is shown. DETAILED DESCRIPTION

[0023] The application proposes a news event detection method based on multi-dimensional heterogeneous graph representation and matching. First, a news data representation and embedding method based on multi-dimensional heterogeneous graph is proposed, which solves the unified modeling and embedding problem of multiple attributes of news data; second, the news multi-dimensional heterogeneous graph is matched and calculated through the graph neural network, which solves the problem of difficult matching between graph structures and news similarity calculation; finally, the dynamic graph model is used to detect news events, which solves the problems that the graph-based news event detection cannot model the dynamic changes of events over time and cannot identify related events. Through the application, the efficiency of news event detection can be effectively improved, and news events can be accurately detected in different fields.

[0024] Specifically, the news event detection method based on multi-dimensional heterogeneous graph representation and matching includes the following steps:

[0025] (1) Extract the main features of news data, including theme features, time features, geographic

[0026] position features, and regard them as node information to construct a news multi-dimensional heterogeneous graph, and input part is a training data set;

[0027]

[0028] (2) Learn the embedding value of the established news multi-dimensional heterogeneous graph, and finally generate the optimal model parameters through step-by-step training;

[0029]

[0030] (3) Use the graph neural network to match the constructed news multi-dimensional heterogeneous graph and calculate the matching score between news;

[0031] (4) Dynamically detect news events, first create a dynamic graph model for news data, and then use the subgraph covered by the sliding time window in the time interval to independently detect events. ​​

[0032] The application takes into account various information such as the theme, time, geographical location, etc. of news, fully utilizes the key information in the news, and establishes the association relationship between various key information, and has stronger news representation ability than the traditional method. The proposed news multi-dimensional heterogeneous graph representation matching method uses a heterogeneous graph embedding technology to obtain embedding values of nodes in the heterogeneous graph, and then encodes the node pairs in the graph through a matching network to obtain matching vectors, and then uses a graph convolution network to aggregate the local matching vectors into the final matching score of the two news, thereby solving the problem that the graph model is difficult to match in a non-Euclidean space, and more accurately comparing the similarity between news. The proposed news event detection method based on a dynamic graph model divides the news multi-dimensional heterogeneous graph sequence using a sliding time window Δt, and independently detects events in the heterogeneous graph sequence in each time window, which can more accurately detect news events and identify the association relationship between events. The news event detection efficiency can be effectively improved. The method proposed in the application can be used for news event detection and evolution, so that the regulatory department has a clearer understanding of the development law and evolution mode of the event, and can more accurately predict the trend of the event development. It has important significance for monitoring focus news and public opinion management in the context of massive data in today's society.

[0033] Preferably, as shown in Figure 2 Step (1) comprises the following sub-steps:

[0034] (1.1) extracting the theme feature of news, using an LDA theme model to extract the theme of each news text, obtaining a theme set T={t1, t2, …t n} for each theme t i ∈T, using a Bert pre-training model to represent it as a k-dimensional vector, and selecting the most probable theme as the theme feature of the news for each news;

[0035] (1.2) extracting the time feature of news, dividing the news publishing time in the data set into k time segments in chronological order, each time segment is of fixed length, and each segment is assigned a unique index in [1, k], for a news, using a hash index to map it to the corresponding time segment according to its time stamp, and using one-hot to encode its time feature into a k-dimensional vector;

[0036] (1.3) extracting the geographical location feature of news, using a named entity recognition method BiLSTM+CRF to identify the geographical location name in each news, and using a Bert pre-training model to represent the geographical location name as a k-dimensional word vector as the geographical location feature;

[0037] (1.4) Construct a news multi-dimensional heterogeneous graph, treat various features as nodes of the multi-dimensional heterogeneous graph, take the news event label as the center, and use the theme feature, time feature, and geographic location feature as the associated nodes to establish a news multi-dimensional heterogeneous graph.

[0038] Preferably, if Figure 3 As shown, the step (2) includes the following sub-steps:

[0039] (2.1) Mapping several different types of node features in the news heterogeneous graph into the same latent vector space. The mapping process uses linear transformation to convert the original feature vector into the latent vector space.

[0040] In vector, realize the conversion of node vector;

[0041] (2.2) Formulate the meta-path P of the multi-dimensional heterogeneous graph, mark the news event label as C, the time node as T, the geographical location node as L, and the topic node as H, establish three types of edges for the multi-dimensional heterogeneous graph, #CT, #CL and #CH, and formulate nine different types of meta-paths, namely CLC, CHC, CTC, TCLCT, TCHCT, LCTCL, LCHCL, HCTCH, and HCLCH;

[0042] (2.3) Aggregate the formulated meta-paths to obtain the node vectors on the meta-paths. For a meta-path P, the intra-meta-path aggregation layer learns the structural and semantic information embedded in the target node, the meta-path-based neighbors, and their contexts by encoding the meta-path instances of P. The meta-path instance is an instance connecting the target node v and the meta-path-based neighbors u. The intra-meta-path aggregation uses a relational rotation encoder to convert all node features on the meta-path instance into a single vector. On this basis, the inter-meta-path aggregation layer is used to integrate the semantic information revealed by the meta-paths. Assume that for node type A, node v∈V A , there is a set of latent vectors: Where m is the number of meta-paths, and the meta-path aggregation uses the attention mechanism to assign different weights to different meta-paths and A The node vectors on the converted meta-path are averaged, and the feature vectors of the meta-path of v are fused using the attention mechanism to obtain the node vector on the meta-path;

[0043] (2.4) The heterogeneous graph embedding module is trained. The relational rotation encoder is used as the encoder of the meta-path instance to model the sequential structural information in the meta-path. The node vectors on the heterogeneous graph meta-path are trained using semi-supervised learning. A part of the labeled nodes is used to optimize the model weights by minimizing the cross entropy through back propagation and gradient descent, thereby learning the node embedding in the heterogeneous graph.

[0044] Preferably, as shown in Figure 4 Step (3) comprises the following sub-steps:

[0045] (3.1) Learning the matching vector of the news multi-dimensional heterogeneous graph node to represent the similarity of two nodes, using a twin network, embedding the node value of the heterogeneous graph as input, and encoding the node embedding into two context vectors through the context layer sharing the same weight, to capture the context information in the news, the context layer contains multiple bidirectional LSTM, and then using an aggregation layer to aggregate the local matching results to obtain the final matching vector, the aggregation layer connects the element level difference and element level product of the two context vectors;

[0046] (3.2) Using a graph convolutional network to aggregate the local matching vector into the final matching score of the two news, using GCN to capture the content in the graph, the input is the adjacency matrix composed of heterogeneous graph matching vectors, the value in the matrix represents the similarity between nodes, through the calculation of the graph convolutional network, the hidden representation of the last layer of the graph convolutional network is merged into a fixed length vector, which is obtained by taking the average value of the hidden vectors of all nodes in the last layer, and the final matching score is calculated by a multilayer perceptron to represent the relationship between the two news.

[0047] Preferably, as shown in Figure 5 Step (4) comprises the following sub-steps:

[0048] (4.1) Building a dynamic graph model, arranging the news multi-dimensional heterogeneous graph according to the time sequence of news publishing to form a heterogeneous graph sequence G={g1,g2,…g n t}, and dividing the graph sequence using a sliding time window Δt to complete the creation of the dynamic graph model;

[0049] (4.2) Detecting events in the news data set, for the built dynamic graph model, on the multi-dimensional heterogeneous graph subset G (Δ t) covered by the sliding time window Δt, using the k-means clustering algorithm to cluster the heterogeneous graph sequence, and using the matching score of the news as the similarity of the news in the clustering process;

[0050] (4.3) Given a set of news events in a time window t and the news events detected in the previous time window t-1 The goal of related event identification is to determine the relevance of event e t ∈E t and event e t-1 ∈E t-1 , the relevance of event e t ∈E ta set of keywords of the first event and a set of keywords of the second event t-1 ∈E t-1 The similarity comparison is performed between the set of keywords of the first event and the set of keywords of the second event to determine the relevance of the events. If the similarity is higher than a set threshold, it is considered that the two events are relevant, otherwise, they are not relevant.

[0051] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment methods can be completed by programs instructing related hardware. The programs can be stored in a computer-readable storage medium. When the programs are executed, each step of the above-mentioned embodiment methods is included, and the storage medium can be ROM / RAM, a magnetic disc, an optical disc, a memory card, etc. Therefore, corresponding to the method of the present application, the present application also simultaneously includes a news event detection device based on multi-dimensional heterogeneous graph representation and matching. The device is usually represented in the form of a functional module corresponding to each step of the method. As shown in the following figure, the device includes: Figure 1

[0052] a news multi-dimensional heterogeneous graph modeling module configured to extract main features of news data, including theme features, time features, and geographic location features, and to construct a news multi-dimensional heterogeneous graph by taking the features as node information, and inputting part of the training data set;

[0053] a news multi-dimensional heterogeneous graph embedding training module configured to learn embedding values of the established news multi-dimensional heterogeneous graph, and finally generate optimal model parameters through step-by-step iterative training;

[0054] a news multi-dimensional heterogeneous graph representation matching module configured to use a graph neural network to match the constructed news multi-dimensional heterogeneous graph and calculate the matching score between news;

[0055] a news event detection module based on a dynamic graph model, configured to dynamically detect news events. First, a dynamic graph model is created for news data, and then subgraphs in a time interval covered by a sliding time window are used to independently detect events.

[0056] Preferably, the news multi-dimensional heterogeneous graph modeling module includes:

[0057] a theme feature extraction unit configured to extract theme features of news, use an LDA theme model to extract themes from each news text, and obtain a theme set T = {t1, t2, … t n} for each theme t i ∈T, use a Bert pre-training model to represent it as a k-dimensional vector, and for each news, select the most probable theme as the theme feature of the news;

[0058] ​The time feature extraction unit extracts the time feature of the news, divides the news publishing time in the data set into k time segments in chronological order, each time segment is of fixed length, and each segment is assigned a unique index in [1, k]. For a news, according to its timestamp, it is mapped to the corresponding time segment using hash index, and its time feature is encoded into a k-dimensional vector using one-hot;

[0059] The geographical position feature extraction unit extracts the geographical position feature of the news. The named entity recognition method BiLSTM+CRF is used to identify the geographical position name in each news, and the Bert pre-training model is used to represent the geographical position name as a k-dimensional word vector as the geographical position feature.

[0060] The multi-dimensional heterogeneous graph modeling unit constructs a multi-dimensional heterogeneous graph of news. Multiple features are regarded as nodes of the multi-dimensional heterogeneous graph, the news event label is taken as the center, and the theme feature, time feature and geographical position feature are taken as associated nodes to establish a multi-dimensional heterogeneous graph of news.

[0061] Preferably, the news multi-dimensional heterogeneous graph embedding training module comprises:

[0062] The node content conversion unit maps several different types of node features in the news heterogeneous graph to the same latent vector space. The mapping process uses linear transformation to convert the original feature vector to the latent vector, realizing the conversion of the node vector.

[0063] The meta-path specification unit specifies the meta-path P of the multi-dimensional heterogeneous graph. The news event label is marked as C, the time node is marked as T, the geographical position node is marked as L, and the theme node is marked as H. Three types of edges #C-T, #C-L and #C-H are established for the multi-dimensional heterogeneous graph, and nine different types of meta-paths are specified, which are CLC, CHC, CTC, TCLCT, TCHCT, LCTCL, LCHCL, HCTCH and HCLCH.

[0064] The meta-path aggregation unit aggregates the specified meta-paths to obtain the node vectors on the meta-paths. For the meta-path P, the meta-path intra-aggregation layer learns the structural and semantic information between the target node, the meta-path-based neighbor, and their context by encoding the meta-path instance of P, wherein the meta-path instance is an instance connecting the target node v and the meta-path-based neighbor u. The meta-path intra-aggregation uses a relation rotation encoder to convert all node features on the meta-path instance into a single vector. On this basis, the meta-path inter-aggregation layer integrates the semantic information revealed by the meta-path. Assuming that for node type A, node v∈V A , there is a latent vector set: wherein m is the number of meta-paths, the inter-meta-path aggregation assigns different weights to different meta-paths by using an attention mechanism, and the feature vectors of all nodes v∈V A The node vectors on the converted meta-path are averaged, the feature vectors of the meta-path of v are fused using an attention mechanism, and the node vectors on the meta-path are obtained.

[0065] The training unit trains the heterogeneous graph embedding module, uses a relation rotation encoder as an encoder of the meta-path instance to model the sequential structure information in the meta-path, trains the node vectors on the meta-path of the heterogeneous graph by using semi-supervised learning, adopts a part of labeled nodes, and optimizes the model weights by minimizing the cross-entropy through back propagation and gradient descent, so as to learn the node embedding in the heterogeneous graph.

[0066] Preferably, the news multi-dimensional heterogeneous graph representation matching module comprises:

[0067] The matching vector encoding unit learns the matching vector of the node of the news multi-dimensional heterogeneous graph to represent the similarity of two nodes, uses a twin network to take the node embedding value of the heterogeneous graph as input, and encodes the node embedding into two context vectors through a context layer with the same weight sharing, which is used to capture the context information in the news, the context layer contains multiple bidirectional LSTMs, and then an aggregation layer is used to aggregate the local matching results to obtain the final matching vector, the aggregation layer connects the element level difference and the element level product of the two context vectors;

[0068] The matching vector aggregation unit aggregates the local matching vectors into the final matching score of the two news by using a graph convolution network, uses the GCN to capture the content in the graph, the input of the GCN is an adjacency matrix composed of the heterogeneous graph matching vectors, the value in the matrix represents the similarity between the nodes, through the calculation of the graph convolution network, the hidden representation of the last layer of the graph convolution network is merged into a fixed-length vector, and the final matching score is obtained by taking the average value of the hidden vectors of all nodes in the last layer, and the final matching score is calculated by a multilayer perceptron to represent the relationship between the two news.

[0069] Preferably, the news event detection module based on the dynamic graph model comprises:

[0070] The dynamic graph model modeling unit constructs a dynamic graph model, arranges the news multi-dimensional heterogeneous graph according to the time sequence of news publishing to form a heterogeneous graph sequence G={g1,g2,…g n}, and divides the graph sequence by using a sliding time window Δt to complete the creation of the dynamic graph model; the event detection unit detects the events in the news data set; for the constructed dynamic graph model, the multi-dimensional heterogeneous graph subset G (Δt)In the above, the k-means clustering algorithm is used to cluster the heterogeneous graph sequence, and in the clustering process, the matching score of the news is used as the similarity of the news;

[0071] The associated event recognition unit is given a set of news events within a given time window t And the news events detected within the previous time window t-1 The goal of the associated event recognition is to determine whether the news event e t ∈E t is associated with e t-1 ∈E t-1 , the keyword set of the event e t ∈E t is compared with the keyword set of e t-1 ∈E t-1 to determine the association of the events, if the similarity is higher than the set threshold, it is considered that the two events are associated, otherwise, they are not associated.

[0072] The above is only a preferred embodiment of the present application, and does not limit the present application in any form, any simple modification, equivalent change and modification of the above embodiment according to the technical essence of the present application are still within the protection scope of the technical scheme of the present application.

Claims

1. A method for detecting news events based on multi-dimensional heterogeneous graph representation and matching, characterized in that: The method comprises the following steps: (1) extracting main features of news data, including theme features, time features, and geographical location features, and taking them as node information to construct a news multidimensional heterogeneous graph, and inputting part as a training data set; (2) learning embedding values of the established news multidimensional heterogeneous graph, and finally generating optimal model parameters through step-by-step iterative training; (3) using a graph neural network to match the constructed news multidimensional heterogeneous graph and calculate the matching score between news; (4) dynamically detecting news events, first creating a dynamic graph model for news data, and then independently detecting events using subgraphs in the time interval covered by a sliding time window; The step (1) comprises the following sub-steps: (1.1) Extract the theme features of the news, use the LDA theme model to extract the theme of each news text, get the theme set T={t1, t2, …t n},for each theme t i ∈T, use the Bert pre-training model to represent it as a k-dimensional vector, and select the theme with the highest probability as the theme feature of the news for each news. (1.2) extracting the time features of news, dividing the news publishing time in the data set into k time segments in chronological order, each time segment being of fixed length, and assigning each segment a unique index in [1, k], and for a news, using hash index to map it to the corresponding time segment according to its timestamp, and using one-hot to encode its time feature into a k-dimensional vector; (1.3) extracting the geographical location features of news, using the named entity recognition method BiLSTM+CRF to identify the geographical location names in each news, and using the Bert pre-training model to represent the geographical location names as k-dimensional word vectors as geographical location features; (1.4) constructing a news multidimensional heterogeneous graph, taking multiple features as nodes of a multidimensional heterogeneous graph, taking news event labels as the center, and taking theme features, time features, and geographical location features as associated nodes to establish a news multidimensional heterogeneous graph. 2.The method of claim 1, wherein the method further comprises: determining a plurality of news events based on the multi-dimensional heterogeneous graph representation and the matching. The step (2) comprises the following sub-steps: (2.1) mapping several different types of node features in the news heterogeneous graph to the same latent vector space, and using linear transformation to convert the original feature vector to the latent vector to realize the conversion of the node vector; (2.2) formulating a meta-path P of the multidimensional heterogeneous graph, marking news event labels as C, time nodes as T, geographical location nodes as L, and theme nodes as H, establishing three types of edges #C-T, #C-L and #C-H for the multidimensional heterogeneous graph, and formulating nine different types of meta-paths, namely CLC, CHC, CTC, TCLCT, TCHCT, LCTCL, LCHCL, HCTCH and HCLCH; (2.3) Aggregate the formulated meta-paths to obtain the node vectors on the meta-paths. For a meta-path P, the intra-meta-path aggregation layer learns the structural and semantic information embedded in the target node, the meta-path-based neighbors, and their contexts by encoding the meta-path instances of P. The meta-path instance is an instance connecting the target node v and the meta-path-based neighbors u. The intra-meta-path aggregation uses a relational rotation encoder to convert all node features on the meta-path instance into a single vector. On this basis, the inter-meta-path aggregation layer is used to integrate the semantic information revealed by the meta-paths. Assume that for node type A, node v∈V A , there is a set of latent vectors: Where m is the number of meta-paths, and the meta-path aggregation uses the attention mechanism to assign different weights to different meta-paths and A The node vectors on the converted meta-path are averaged, and the feature vectors of the meta-path of v are fused using the attention mechanism to obtain the node vector on the meta-path; (2.4) training the heterogeneous graph embedding module, using a relation rotation encoder as an encoder of the meta-path instance to model the sequential structure information in the meta-path, training the node vector on the meta-path of the heterogeneous graph using semi-supervised learning, adopting a part of labeled nodes to minimize the cross-entropy of the model weight through back propagation and gradient descent to optimize the model weight, and thus learning the node embedding in the heterogeneous graph. 3.The method of claim 2, wherein the method further comprises: determining a plurality of news events based on the multi-dimensional heterogeneous graph representation and the matching. The step (3) comprises the following sub-steps: (3.1) Learning the matching vector of the news multi-dimensional heterogeneous graph node to represent the similarity between two nodes, using a twin network, embedding the value of the node of the heterogeneous graph as input, and encoding the node embedding into two context vectors through the context layer with the same weight, to capture the context information in the news, the context layer contains multiple bidirectional LSTM, and then using the aggregation layer to aggregate the local matching results to obtain the final matching vector, the aggregation layer connects the element level difference and element level product of the two context vectors; (3.2) Using graph convolution network to aggregate the local matching vector into the final matching score of the two news, using GCN to capture the content in the graph, the input is the adjacency matrix composed of heterogeneous graph matching vectors, the value in the matrix represents the similarity between nodes, through the calculation of graph convolution network, the hidden representation of the last layer of graph convolution network is merged into a fixed length vector, the final matching score is obtained by taking the average value of the hidden vectors of all nodes in the last layer, and the final matching score is calculated by multilayer perception to represent the relationship between the two news. 4.The method of claim 3, wherein the method further comprises: determining a plurality of news events based on the multi-dimensional heterogeneous graph representation and the matching. The step (4) comprises the following sub-steps: (4.1) Construct a dynamic graph model, arrange the news multi-dimensional heterogeneous graph according to the time sequence of news release, form a heterogeneous graph sequence G={g1, g2, …g n} and divide the graph sequence using a sliding time window Δt, complete the creation of the dynamic graph model; (4.2) detecting events in the news dataset, for the constructed dynamic graph model, in the multi-dimensional heterogeneous graph subset G covered by the sliding time window Δt (Δt) Above, using the k-means clustering algorithm to cluster the heterogeneous graph sequence, in the clustering process, using the matching score of the news as the similarity of the news; (4.3) a set of news events within a given time window t and news events detected within the previous time window t-1 The goal of the related event identification is to determine the relevance of a news event e t ∈ E t to e t-1 ∈ E t-1 The keyword set of event e t ∈ E t is compared to the keyword set of e t-1 ∈ E t-1 to determine the relevance of the events, and if the similarity is higher than a set threshold, then the two events are considered to be related, otherwise not.

5. The device for detecting news event based on multi-dimensional heterogeneous graph representation and matching, characterized in that: It comprises: The news multi-dimensional heterogeneous graph modeling module is configured to extract the main features of the news data, including the theme feature, the time feature and the geographical location feature, and regards them as node information to construct the news multi-dimensional heterogeneous graph, and the input part is the training data set; The news multi-dimensional heterogeneous graph embedding training module is configured to learn the embedding value of the established news multi-dimensional heterogeneous graph, and finally generate the optimal model parameter through step-by-step iterative training; The news multi-dimensional heterogeneous graph representation matching module is configured to use the graph neural network to match the constructed news multi-dimensional heterogeneous graph and calculate the matching score between news; The news event detection module based on dynamic graph model is configured to dynamically detect news events, first create a dynamic graph model for news data, and then independently detect events using the sub-graph in the time interval covered by the sliding time window; The news multi-dimensional heterogeneous graph modeling module comprises: A subject feature extraction unit extracts subject features of news, uses an LDA subject model to extract a subject of each news text to obtain a subject set T = {t1, t2, … t n} and, for each subject t i ∈T, uses a Bert pre-training model to represent it as a k-dimensional vector, and selects a subject with the highest probability as a subject feature of each news. The time feature extraction unit extracts the time feature of the news, divides the news publishing time in the data set into k time segments in chronological order, each time segment is of fixed length, and each segment is assigned a unique index in [1, k]. For a news, according to its time stamp, it is mapped to the corresponding time segment using hash index, and its time feature is encoded into a k-dimensional vector using one-hot; The geographical location feature extraction unit extracts the geographical location feature of the news, uses the named entity recognition method BiLSTM+CRF to identify the geographical location name in each news, and uses the Bert pre-training model to represent the geographical location name as a k-dimensional word vector as the geographical location feature; The multi-dimensional heterogeneous graph modeling unit constructs the news multi-dimensional heterogeneous graph, regards multiple features as nodes of the multi-dimensional heterogeneous graph, takes the news event label as the center, takes the theme feature, the time feature and the geographical location feature as the associated nodes respectively, and establishes the news multi-dimensional heterogeneous graph.

6. The device according to claim 5, wherein the device is characterized in that: The news multi-dimensional heterogeneous graph embedding training module comprises: A node content conversion unit maps several different types of node features in the news heterogeneous graph into the same latent vector space, and a mapping process uses linear transformation to convert the original feature vector into a latent vector, thereby realizing conversion of the node vector; An element path formulation unit formulates an element path P of the multi-dimensional heterogeneous graph, labels a news event label as C, a time node as T, a geographical location node as L, and a theme node as H, establishes three types of edges #C-T, #C-L and #C-H for the multi-dimensional heterogeneous graph, and formulates nine different types of element paths, namely CLC, CHC, CTC, TCLCT, TCHCT, LCTCL, LCHCL, HCTCH and HCLCH; The meta-path aggregation unit aggregates the formulated meta-path, obtains the node vector on the meta-path, and for the meta-path P, the intra-meta-path aggregation layer learns the structural and semantic information embedded between the target node, the meta-path-based neighbor, and their context by encoding the meta-path instance of P, wherein the meta-path instance is an instance connecting the target node v and the meta-path-based neighbor u, the intra-meta-path aggregation adopts a relation rotation encoder to convert all node features on the meta-path instance into a single vector, and on this basis, the inter-meta-path aggregation layer integrates the semantic information revealed by the meta-path, and it is assumed that for the node type A, the node v is in V A , there is a set of latent vectors: wherein m is the number of meta-paths, the inter-meta-path aggregation uses an attention mechanism to assign different weights to different meta-paths, and for all nodes v in V A The average value of the converted node vector on the meta-path is obtained, the feature vectors of the meta-paths of v are fused using an attention mechanism, and the node vector on the meta-path is obtained. A training unit trains the heterogeneous graph embedding module, uses a relation rotation encoder as an encoder of the element path instance, models the sequential structure information in the element path, trains the node vector on the element path of the heterogeneous graph by using semi-supervised learning, adopts a part of the labeled nodes, and minimizes the cross-entropy by using back propagation and gradient descent to optimize the model weight, thereby learning the node embedding in the heterogeneous graph.

7. The device for detecting news events based on multi-dimensional heterogeneous graph representation and matching according to claim 6, characterized in that: The news multi-dimensional heterogeneous graph representation matching module comprises: A matching vector encoding unit learns a matching vector of a node of the news multi-dimensional heterogeneous graph to represent the similarity between two nodes, uses a twin network, takes the node embedding value of the heterogeneous graph as input, and encodes the node embedding into two context vectors by using a context layer with the same weight, so as to capture the context information in the news, the context layer comprises a plurality of bidirectional LSTM, and then an aggregation layer is used to aggregate the local matching result to obtain a final matching vector, the aggregation layer is connected with the element level difference and the element level product of the two context vectors; 8.The news event detection device based on multi-dimensional heterogeneous graph representation and matching according to claim 7, characterized in that: A matching vector aggregation unit uses a graph convolution network to aggregate the local matching vector into a final matching score of two news, uses the GCN to capture the content in the graph, the input of the GCN is an adjacency matrix composed of the heterogeneous graph matching vector, the value in the matrix represents the similarity between nodes, through calculation of the graph convolution network, the hidden representation of the last layer of the graph convolution network is merged into a fixed-length vector, the final matching score is obtained by taking the average value of the hidden vectors of all nodes in the last layer, and the final matching score is calculated by using a multilayer perception to represent the relationship between the two news. The news event detection module based on the dynamic graph model comprises: Dynamic graph modeling unit, constructs a dynamic graph model, arranges the news multi-dimensional heterogeneous graph according to the time sequence of news release, and forms a heterogeneous graph sequence G = {g1, g2, ... g n }, and use the sliding time window Δt to divide the graph sequence to complete the creation of the dynamic graph model; the event detection unit detects events in the news dataset; for the constructed dynamic graph model, the multi-dimensional heterogeneous graph subset G covered by the sliding time window Δt (Δt) In the above, the k-means clustering algorithm is used to cluster heterogeneous graph sequences. In the clustering process, the matching score of the news is used as the similarity of the news; The correlation event recognition unit detects a set of news events within a given time window t and news events detected within a previous time window t-1 The goal of the related event identification is to determine the relevance of news events e t ∈ E t to e t-1 ∈ E t-1 . The keyword set of e t ∈ E t is compared with the keyword set of e t-1 ∈ E t-1 to determine the relevance of the events, and if the similarity is higher than a set threshold, then the two events are considered to be related, otherwise not.

Images