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

Abstract

The invention discloses a news event detection method and device based on multi-dimensional heterogeneous graph representation and matching, and solves the problems of unified modeling and embedding of various attributes of news data, difficult matching between graph structures and news similarity calculation, and the problem that graph-based news event detection cannot model, events dynamically change along with time, and the news event detection efficiency is improved. And the problem that the news event detection efficiency based on the graph is low is solved, and the news event can be accurately detected in different fields. The method comprises the following steps: (1) extracting main features of news data, regarding the main features as node information, and constructing a news multi-dimensional heterogeneous graph; (2) learning the news multi-dimensional heterogeneous graph embedded value, and carrying out iterative training step by step to generate an optimal model parameter; (3) matching the multi-dimensional heterogeneous graphs of the news by using a graph neural network, and calculating a matching score between the news; and (4) creating a dynamic graph model for the news data, and dynamically detecting a news event by using a sub-graph in a time interval covered by a sliding time window.

Description

Technical Field

[0001] The present invention 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 Art

[0002] With the rapid development of the Internet and various news media platforms, people's access to information is becoming increasingly diverse. People can get the latest news and information from all over the world anytime and anywhere through smartphones and tablets, whether it is international political dynamics, economic development, technological innovation breakthroughs or popular culture trends. The massive news resources provided by these platforms have greatly enriched and facilitated people's daily lives. However, while bringing convenience to people's lives, there are also many challenges. The most prominent challenge is the increasingly serious phenomenon of information overload. People often find it difficult to quickly identify valuable content in the vast flood of information. This information overload not only consumes a lot of users' time and energy, but also may lead to misreading and misunderstanding of information, making it more difficult to distinguish the truth and grasp the key points. This has become a problem that users generally pay attention to. News event detection is an effective way to solve this problem. It uses advanced computer algorithms and big data analysis technology to automatically identify, classify and track the occurrence and development of news events. Through real-time monitoring and intelligent analysis of large amounts of news data, this technology can help regulatory authorities, media organizations and even ordinary users quickly identify hot events, thereby effectively alleviating the pressure caused by information overload. News event detection technology can not only improve the efficiency of information processing, but also help build a healthier and more orderly network information environment.

[0003] At present, the research on news event detection includes traditional feature-based methods, topic model-based methods, and graph-based methods. Among them, graph-based methods are the most popular news event detection methods. However, since most graph-based event detection methods rely heavily on linguistic knowledge and the semantic features of news content, they bring about problems such as information loss, semantic sparsity, and inaccurate event detection. In addition, this type of method requires the use of community detection algorithms to divide the graph, but due to the high time complexity of community detection algorithms, graph-based event detection methods have a large time overhead under large-scale data, resulting in limited application scope. Summary of the invention

[0004] In order to overcome the defects of the prior art, the technical problem to be solved by the present invention is to provide a news event detection method based on multi-dimensional heterogeneous graph representation and matching, which solves the problem of unified modeling and embedding of multiple attributes of news data, solves the problem of difficulty in matching between graph structures and calculating the similarity of news, solves the problem that graph-based news event detection cannot model the dynamic changes of events over time and cannot identify related events, solves the problem of low efficiency of graph-based news event detection, can accurately detect news events in different fields, and effectively improve the efficiency of news event detection.

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

[0006] (1) Extract the main features of news data, including topic features, time features, and geographic location features, and regard them as node information to construct a news multi-dimensional heterogeneous graph, with the input part being 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) Use graph neural networks to match the constructed news multi-dimensional heterogeneous graph and calculate the matching scores between news;

[0009] (4) Dynamically detect news events. First, create a dynamic graph model for news data.

[0010] The event is then detected independently on the subgraphs within the time interval covered by the sliding time window.

[0011] The present invention takes into account various information such as the subject, time, and geographic location of the news, makes full use of the key information in the news, and establishes the correlation between multiple key information, which has stronger news representation capabilities than traditional methods. The proposed news multi-dimensional heterogeneous graph representation matching method uses heterogeneous graph embedding technology to obtain the embedding value of the heterogeneous graph node. On this basis, the node pairs in the graph are encoded by the matching network to obtain the matching vector, and then the local matching vector is aggregated into the final matching score of the two news using the graph convolution network, which solves the problem that the graph model is difficult to match in non-Euclidean space, and more accurately compares the similarity between news. The proposed news event detection method based on the dynamic graph model uses a sliding time window Δt to divide the news multi-dimensional heterogeneous graph sequence, and independently performs event detection on the heterogeneous graph sequence in each time window, which can more accurately detect news events and identify the correlation between events. It can effectively improve the efficiency of news event detection. The method proposed by the present invention can be used for news event detection and evolution, so that the regulatory authorities have a clearer understanding of the development law and evolution pattern of the event, and can more accurately predict the trend of event development. In today's society with massive data, monitoring hot 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;

[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 sub-graphs within the time interval covered by the sliding time window are used to independently detect events. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0019] Figure 2 A schematic diagram of a news data modeling model based on a multi-dimensional heterogeneous graph according to the present invention is shown.

[0020] Figure 3 A flowchart of news multi-dimensional heterogeneous graph embedding training according to the present invention is shown.

[0021] Figure 4 A flowchart of news representation matching based on a multi-dimensional heterogeneous graph according to the present invention is shown.

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

[0023] The present invention 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 a multi-dimensional heterogeneous graph is proposed, which solves the problem of unified modeling and embedding of multiple attributes of news data; secondly, a matching calculation is performed on the news multi-dimensional heterogeneous graph through a graph neural network, which solves the problem of difficulty in matching between graph structures and similarity calculation of news; finally, a dynamic graph model is used to detect news events, which solves the problem that graph-based news event detection cannot model dynamic changes of events over time and cannot identify related events. Through the present invention, 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 topic features, time features, geographic features,

[0026] The location features are regarded as node information to construct a news multi-dimensional heterogeneous graph.

[0027] The input part is the training data set;

[0028] (2) Learn the embedding value of the established news multi-dimensional heterogeneous graph, and gradually iterate

[0029] Repeated training to finally generate the optimal model parameters;

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

[0031] (4) To dynamically detect news events, we first create a dynamic graph model for the news data, and then use the subgraphs within the time interval covered by the sliding time window to independently detect events.

[0032] The present invention takes into account various information such as the subject, time, and geographic location of the news, makes full use of the key information in the news, and establishes the correlation between multiple key information, which has stronger news representation capabilities than traditional methods. The proposed news multi-dimensional heterogeneous graph representation matching method uses heterogeneous graph embedding technology to obtain the embedding value of the heterogeneous graph node. On this basis, the node pairs in the graph are encoded by the matching network to obtain the matching vector, and then the local matching vector is aggregated into the final matching score of the two news using the graph convolution network, which solves the problem that the graph model is difficult to match in non-Euclidean space, and more accurately compares the similarity between news. The proposed news event detection method based on the dynamic graph model uses a sliding time window Δt to divide the news multi-dimensional heterogeneous graph sequence, and independently performs event detection on the heterogeneous graph sequence in each time window, which can more accurately detect news events and identify the correlation between events. It can effectively improve the efficiency of news event detection. The method proposed by the present invention can be used for news event detection and evolution, so that the regulatory authorities have a clearer understanding of the development law and evolution pattern of the event, and can more accurately predict the trend of event development. In today's society with massive data, monitoring hot news and public opinion management are of great significance.

[0033] Preferably, if Figure 2 As shown, the step (1) includes the following sub-steps:

[0034] (1.1) Extract the topic features of the news and use the LDA topic model to extract topics for each news text, and obtain the topic set T = {t 1 , t 2 ,…t n}, for each topic t i ∈T, use the Bert pre-training model to represent it as a k-dimensional vector. For each news article, select the topic with the highest probability as the topic feature of the news;

[0035] (1.2) Extract the time features of news. Divide the news release time in the data set into k time segments in chronological order. Each time segment is of fixed length. Assign each segment a unique index in [1, k]. For a piece of news, use a hash index to map it to the corresponding time segment based on its timestamp, and use one-hot to encode its time features into a k-dimensional vector.

[0036] (1.3) Extract the geographical location features of the news. Use the named entity recognition method BiLSTM+CRF to identify the geographical location name in each news item, and use the Bert pre-trained 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, regard 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) Map the node features of different types 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 transformation 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 latent vector set: 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) Train the heterogeneous graph embedding module, use the relational rotation encoder as the encoder of the meta-path instance, model the sequential structural information in the meta-path, use semi-supervised learning to train the node vectors on the heterogeneous graph meta-path, use a part of labeled nodes, and optimize the model weights by back propagation and gradient descent to minimize the cross entropy, so as to learn the node embedding in the heterogeneous graph.

[0044] Preferably, if Figure 4 As shown, the step (3) includes the following sub-steps:

[0045] (3.1) Learn the matching vectors of the nodes of the news multi-dimensional heterogeneous graph to represent the similarity between two nodes. Use the Siamese network to take the node embedding values ​​of the heterogeneous graph as input, and encode the node embedding into two context vectors through the context layer that shares the same weight to capture the context information in the news. The context layer contains multiple bidirectional LSTMs, and then use 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.

[0046] (3.2) Use a graph convolutional network to aggregate local matching vectors into the final matching score of the two news articles. Use GCN to capture the content in the graph. Its input is the adjacency matrix composed of heterogeneous graph matching vectors. The values ​​in the matrix represent 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 vector of fixed length, which is obtained by taking the average of the hidden vectors of all nodes in the last layer. The final matching score will be calculated by a multi-layer perceptron to represent the relationship between the two news articles.

[0047] Preferably, if Figure 5 As shown, the step (4) includes the following sub-steps:

[0048] (4.1) Construct a dynamic graph model and arrange the news multi-dimensional heterogeneous graph in the chronological order of news release to form a heterogeneous graph sequence G = {g 1 ,g 2 ,…g n}, and use the sliding time window Δt to divide the graph sequence to complete the creation of the dynamic graph model;

[0049] (4.2) Detecting 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 (Δ t), the k-means clustering algorithm is used to cluster the heterogeneous graph sequences. In the clustering process, the matching score of the news is used as the similarity of the news;

[0050] (4.3) A set of news events within a given time window t and the news events detected in the previous time window t-1 The goal of related event recognition is to determine the news events t ∈E t With e t-1 ∈E t-1The relevance of event e t ∈E t Keyword collection and e t-1 ∈E t-1 The keyword sets are compared for similarity to determine the relevance of events. If the similarity is higher than the set threshold, it is considered that the two events are related, otherwise there is no correlation.

[0051] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be implemented by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes the steps of the above-mentioned embodiment method, and the storage medium can be: ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a news event detection device based on multi-dimensional heterogeneous graph representation and matching, which is usually represented in the form of functional modules corresponding to the steps of the method. Figure 1 As shown, the device comprises:

[0052] 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 to construct a news multi-dimensional heterogeneous graph. The input part is the training data set;

[0053] 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;

[0054] 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;

[0055] 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 sub-graphs within the time interval covered by the sliding time window are used to independently detect events.

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

[0057] The topic feature extraction unit extracts the topic features of the news and uses the LDA topic model to extract topics for each news text to obtain the topic set T = {t 1 , t 2 ,…t n}, for each topic t i∈T, use the Bert pre-training model to represent it as a k-dimensional vector. For each news article, select the topic with the highest probability as the topic feature of the news;

[0058] The time feature extraction unit extracts the time features of news. The news release time in the data set is divided into k time segments in chronological order. Each time segment is of fixed length. Each segment is assigned a unique index in [1, k]. For a piece of news, a hash index is used to map it to the corresponding time segment according to its timestamp, and its time feature is encoded into a k-dimensional vector using one-hot.

[0059] The geographic location feature extraction unit extracts the geographic location features of the news. It uses the named entity recognition method BiLSTM+CRF to identify the geographic location name in each news item, and uses the Bert pre-trained model to represent the geographic location name as a k-dimensional word vector as the geographic location feature.

[0060] The multi-dimensional heterogeneous graph modeling unit constructs a news multi-dimensional heterogeneous graph, regards multiple features as nodes of the multi-dimensional heterogeneous graph, takes news event labels as the center, and uses topic features, time features, and geographic location features as associated nodes to establish a news multi-dimensional heterogeneous graph.

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

[0062] The node content conversion unit maps 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 a latent vector to achieve the conversion of the node vector.

[0063] The meta-path formulation unit formulates the meta-path P of the multi-dimensional heterogeneous graph, marks the news event label as C, the time node as T, the geographical location node as L, and the topic node as H, establishes three types of edges for the multi-dimensional heterogeneous graph, #CT, #CL and #CH, and formulates nine different types of meta-paths, namely CLC, CHC, CTC, TCLCT, TCHCT, LCTCL, LCHCL, HCTCH, and HCLCH;

[0064] The meta-path aggregation unit aggregates 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 latent vector set: 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;

[0065] The training unit trains the heterogeneous graph embedding module, uses the relational rotation encoder as the encoder of the meta-path instance, models the sequential structural information in the meta-path, uses semi-supervised learning to train the node vectors on the heterogeneous graph meta-path, uses a part of labeled nodes, and optimizes the model weights by back propagation and gradient descent to minimize the cross entropy, thereby learning the node embedding in the heterogeneous graph.

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

[0067] The matching vector encoding unit learns the matching vectors of the nodes in the news multi-dimensional heterogeneous graph to represent the similarity between two nodes. It uses the twin network to take the node embedding values ​​of the heterogeneous graph as input and encodes the node embedding into two context vectors through the context layer that shares the same weight to capture the context information in the news. The context layer contains multiple bidirectional LSTMs, and then uses 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.

[0068] The matching vector aggregation unit uses a graph convolutional network to aggregate local matching vectors into the final matching score of two news articles. It uses GCN to capture the content in the graph. Its input is the adjacency matrix composed of heterogeneous graph matching vectors. The values ​​in the matrix represent 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 vector of fixed length, which is obtained by taking the average value of the hidden vectors of all nodes in the last layer. The final matching score will be calculated by a multi-layer perceptron to represent the relationship between the two news articles.

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

[0070] The dynamic graph model modeling unit constructs a dynamic graph model and arranges the news multi-dimensional heterogeneous graph in the chronological order of news release to form a heterogeneous graph sequence G = {g 1 ,g 2 ,…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 the heterogeneous graph sequence. In the clustering process, the matching score of the news is used as the similarity of the news;

[0071] The associated event recognition unit, a set of news events within a given time window t and the news events detected in the previous time window t-1 The goal of related event recognition is to determine the news events t ∈E t With e t-1 ∈E t-1 The relevance of event e t ∈E t Keyword collection and e t-1 ∈E t-1 The keyword sets are compared for similarity to determine the relevance of events. If the similarity is higher than the set threshold, it is considered that the two events are related, otherwise there is no correlation.

[0072] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the protection scope of the technical solution of the present invention.

Claims

1. A news event detection method based on multi-dimensional heterogeneous graph representation and matching, characterized by: The method comprises the following steps: (1) Extract the main features of news data, including topic features, time features, and geographic location features, and regard them as node information to construct a news multi-dimensional heterogeneous graph, with the input part being the training data set; (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; (3) Use graph neural networks to match the constructed news multi-dimensional heterogeneous graph and calculate the matching scores between news; (4) To dynamically detect news events, we first create a dynamic graph model for the news data, and then use the subgraphs within the time interval covered by the sliding time window to independently detect events.

2. The news event detection method based on multi-dimensional heterogeneous graph representation and matching according to claim 1 is characterized in that: The step (1) comprises the following sub-steps: (1.1) Extract the topic features of the news and use the LDA topic model to extract topics for each news text, and obtain the topic set T = {t1, t2, ...t n }, for each topic t i ∈T, use the Bert pre-training model to represent it as a k-dimensional vector. For each news article, select the topic with the highest probability as the topic feature of the news; (1.2) Extract the time features of news. Divide the news release time in the data set into k time segments in chronological order. Each time segment is of fixed length. Assign each segment a unique index in [1, k]. For a piece of news, use a hash index to map it to the corresponding time segment based on its timestamp, and use one-hot to encode its time features into a k-dimensional vector. (1.3) Extract the geographical location features of the news. Use the named entity recognition method BiLSTM+CRF to identify the geographical location name in each news item, and use the Bert pre-trained model to represent the geographical location name as a k-dimensional word vector as the geographical location feature; (1.4) Construct a news multi-dimensional heterogeneous graph, regard 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.

3. The news event detection method based on multi-dimensional heterogeneous graph representation and matching according to claim 2 is characterized in that: The step (2) comprises the following sub-steps: (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 a latent vector to achieve the transformation of the node vector; (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; (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 latent vector set: 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) Train the heterogeneous graph embedding module, use the relational rotation encoder as the encoder of the meta-path instance, model the sequential structural information in the meta-path, use semi-supervised learning to train the node vectors on the heterogeneous graph meta-path, use a part of labeled nodes, and optimize the model weights by back propagation and gradient descent to minimize the cross entropy, so as to learn the node embedding in the heterogeneous graph.

4. The news event detection method based on multi-dimensional heterogeneous graph representation and matching according to claim 3 is characterized in that: The step (3) comprises the following sub-steps: (3.1) Learn the matching vectors of the nodes of the news multi-dimensional heterogeneous graph to represent the similarity between two nodes. Use the Siamese network to take the node embedding values ​​of the heterogeneous graph as input, and encode the node embedding into two context vectors through the context layer that shares the same weight to capture the context information in the news. The context layer contains multiple bidirectional LSTMs, and then use 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) Use a graph convolutional network to aggregate local matching vectors into the final matching score of the two news articles. Use GCN to capture the content in the graph. Its input is the adjacency matrix composed of heterogeneous graph matching vectors. The values ​​in the matrix represent 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 vector of fixed length, which is obtained by taking the average of the hidden vectors of all nodes in the last layer. The final matching score will be calculated by a multi-layer perceptron to represent the relationship between the two news articles.

5. The news event detection method based on multi-dimensional heterogeneous graph representation and matching according to claim 4 is characterized in that: The step (4) comprises the following sub-steps: (4.1) Construct a dynamic graph model and arrange the news multi-dimensional heterogeneous graph in the chronological order of news release to form 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; (4.2) Detecting 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 the heterogeneous graph sequence. In the clustering process, the matching score of the news is used as the similarity of the news; (4.3) A set of news events within a given time window t and the news events detected in the previous time window t-1 The goal of related event recognition is to determine the news events t ∈E t With e t-1 ∈E t-1 The relevance of event e t ∈E t Keyword collection and e t-1 ∈E t-1 The keyword sets are compared for similarity to determine the relevance of events. If the similarity is higher than the set threshold, it is considered that the two events are related, otherwise there is no correlation.

6. A news event detection device based on multi-dimensional heterogeneous graph representation and matching, characterized in that: It includes: 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 to construct a news multi-dimensional heterogeneous graph. 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 parameters through step-by-step iterative training; 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; 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 sub-graphs within the time interval covered by the sliding time window are used to independently detect events.

7. The news event detection device based on multi-dimensional heterogeneous graph representation and matching according to claim 6 is characterized in that: The news multi-dimensional heterogeneous graph modeling module includes: The topic feature extraction unit extracts the topic features of the news and uses the LDA topic model to extract topics for each news text to obtain the topic set T = {t1, t2, ...t n }, for each topic t i ∈T, use the Bert pre-training model to represent it as a k-dimensional vector. For each news article, select the topic with the highest probability as the topic feature of the news; The time feature extraction unit extracts the time features of news. The news release time in the data set is divided into k time segments in chronological order. Each time segment is of fixed length. Each segment is assigned a unique index in [1, k]. For a piece of news, a hash index is used to map it to the corresponding time segment according to its timestamp, and its time feature is encoded into a k-dimensional vector using one-hot. The geographic location feature extraction unit extracts the geographic location features of the news. It uses the named entity recognition method BiLSTM+CRF to identify the geographic location name in each news item, and uses the Bert pre-trained model to represent the geographic location name as a k-dimensional word vector as the geographic location feature. The multi-dimensional heterogeneous graph modeling unit constructs a news multi-dimensional heterogeneous graph, regards multiple features as nodes of the multi-dimensional heterogeneous graph, takes news event labels as the center, and uses topic features, time features, and geographic location features as associated nodes to establish a news multi-dimensional heterogeneous graph.

8. The news event detection device based on multi-dimensional heterogeneous graph representation and matching according to claim 7 is characterized in that: The news multi-dimensional heterogeneous graph embedding training module includes: The node content conversion unit maps 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 a latent vector to achieve the conversion of the node vector. The meta-path formulation unit formulates the meta-path P of the multi-dimensional heterogeneous graph, marks the news event label as C, the time node as T, the geographical location node as L, and the topic node as H, establishes three types of edges for the multi-dimensional heterogeneous graph, #CT, #CL and #CH, and formulates nine different types of meta-paths, namely CLC, CHC, CTC, TCLCT, TCHCT, LCTCL, LCHCL, HCTCH, and HCLCH; The meta-path aggregation unit aggregates 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 latent vector set: 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; The training unit trains the heterogeneous graph embedding module, uses the relational rotation encoder as the encoder of the meta-path instance, models the sequential structural information in the meta-path, uses semi-supervised learning to train the node vectors on the heterogeneous graph meta-path, uses a part of labeled nodes, and optimizes the model weights by back propagation and gradient descent to minimize the cross entropy, thereby learning the node embedding in the heterogeneous graph.

9. The news event detection device based on multi-dimensional heterogeneous graph representation and matching according to claim 8, characterized in that: The news multi-dimensional heterogeneous graph representation matching module includes: The matching vector encoding unit learns the matching vectors of the nodes in the news multi-dimensional heterogeneous graph to represent the similarity between two nodes. It uses the twin network to take the node embedding values ​​of the heterogeneous graph as input and encodes the node embedding into two context vectors through the context layer that shares the same weight to capture the context information in the news. The context layer contains multiple bidirectional LSTMs, and then uses 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. The matching vector aggregation unit uses a graph convolutional network to aggregate local matching vectors into the final matching score of two news articles. It uses GCN to capture the content in the graph. Its input is the adjacency matrix composed of heterogeneous graph matching vectors. The values ​​in the matrix represent 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 vector of fixed length, which is obtained by taking the average value of the hidden vectors of all nodes in the last layer. The final matching score will be calculated by a multi-layer perceptron to represent the relationship between the two news articles.

10. The news event detection device based on multi-dimensional heterogeneous graph representation and matching according to claim 9, characterized in that: The news event detection module based on the dynamic graph model includes: The dynamic graph model modeling unit constructs a dynamic graph model and arranges the news multi-dimensional heterogeneous graph in the chronological order of news release to form 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 the heterogeneous graph sequence. In the clustering process, the matching score of the news is used as the similarity of the news; The associated event recognition unit, a set of news events within a given time window t and the news events detected in the previous time window t-1 The goal of related event recognition is to determine the news events t ∈E t With e t-1 ∈E t-1 The relevance of event e t ∈E t Keyword collection and e t-1 ∈E t-1 The keyword sets are compared for similarity to determine the relevance of events. If the similarity is higher than the set threshold, it is considered that the two events are related, otherwise there is no correlation.

Images