Cross-language news core event analysis method based on graph knowledge distillation

Abstract

The invention relates to a cross-language news core event analysis method based on graph knowledge distillation, and belongs to the technical field of natural language processing. Due to scarcity of news core event analysis annotation data in a low-resource language scene, discourse news event global features and inter-event association features are difficult to mine and model. Aiming at the problem, a cross-language news core event analysis method based on graph knowledge distillation is provided; according to the method, an event hypergraph with event elements as vertexes is constructed to represent event global features, and an event line graph with events as vertexes is constructed to represent correlation features between the events; and event feature knowledge and relation knowledge in a rich resource language are migrated to a low resource language by utilizing cross-language graph knowledge distillation, so that the news core event analysis effect of the low resource language is improved. Experimental results show that the method can effectively mine and model global features of chapter news events and correlation features between the events, and effectively relieve the problem of scarcity of news core event analysis annotation data in a low-resource language scene.

Description

Technical Field

[0001] The present invention relates to a cross-language news core event analysis method based on graph knowledge distillation, and belongs to the technical field of natural language processing. Background Art

[0002] It can quickly and accurately identify key events and related information, providing users with an efficient way to obtain important information. At the same time, this technology has greatly promoted the in-depth analysis and data mining of news content, helping news organizations and researchers to deeply understand the background, impact and relevance of events. In addition, it also supports multilingual and cross-cultural information processing, providing a barrier-free information exchange platform for global users, and improving user experience through personalized content recommendations. At the same time, it also provides news editors and reporters with powerful auxiliary tools, improving work efficiency and reducing repetitive work. Ultimately, the application of news core analysis and detection technology not only improves the level of automation in news processing, but also promotes innovation and development in the entire field of information dissemination.

[0003] News core event analysis aims to analyze the importance of multiple related news events and identify the events that best reflect the core themes of the news. Its core objective is to effectively model the relationships between these events. Due to the scarcity of annotated data for news core event analysis in low-resource language scenarios, mining and modeling the global features of news events and the inter-event correlations are difficult. To address this issue, we propose a cross-language news core event analysis method based on graph knowledge distillation.

[0004] How to effectively model the correlation features between events has become a research hotspot in the past. Researchers have previously constructed methods such as the co-reference relationship of events in news texts, the frequency of event occurrences, and the location of events to alleviate the above problems. However, these methods are still lacking in modeling the global semantic information of events. To overcome this limitation, Jindal et al. tried to improve the performance of core analysis detection by strengthening event representation and encoding potential interactions between events. Wang et al. used a question-answering model to integrate prior knowledge into the event extraction process, and identified core events by integrating similar events and scoring the extracted events. In order to more deeply model the correlation between events, Lu et al. detected core events by constructing complex structural relationships between words and sentences based on the parts of speech and dependency relationships of words in the article.

[0005] However, existing methods mainly focus on monolingual and resource-rich language scenarios, and still face many challenges in cross-lingual and low-resource language scenarios:

[0006] (1) The global semantic information of events cannot be effectively modeled only through discrete event tuples and event global features;

[0007] (2) There are complex correlations between events in news texts, and the attention mechanism alone cannot effectively model the correlation features between events;

[0008] (3) In low-resource scenarios, labeled data is scarce and difficult to construct, resulting in poor model generalization ability. Summary of the Invention

[0009] In order to solve the above problems, the present invention provides a cross-language news core event analysis method based on graph knowledge distillation. The present invention effectively distills rich-resource knowledge to low-resource clients, improving the performance of low-resource language models.

[0010] The present invention proposes a cross-language news core analysis and detection method based on graph knowledge distillation, which effectively distills rich resource knowledge to low resource clients and improves the performance of low resource language models. Specifically, in response to the above-mentioned problems (1) and (2), the local and global semantic information of events are efficiently modeled using hypergraph structures. Specifically, first, based on the internal and relationship between events, an event hypergraph and an event line graph (ELG) are constructed respectively. In the event hypergraph, a single event is regarded as a hyperedge, and its event elements are used as the vertices of the hyperedge. The hyperedges of different events are connected through shared event elements, thereby constructing a complete event hypergraph. In the event line graph, based on the constructed event hypergraph, events are used as vertices, and the connections between line graph vertices are established through shared event elements. Subsequently, in the two channels of the event hypergraph and the event line graph, hypergraph convolution and graph attention network are used respectively to obtain the event embedding of each channel. Finally, cross-view self-supervised learning is introduced to learn the consistency of the event embedding of the two channels, further enhancing the model's ability to capture event context information and model the correlation features between events. To address problem (3), we further proposed a cross-lingual knowledge distillation framework based on hypergraph and line graph modeling. This framework leverages prior knowledge learned in resource-rich languages (English) to compensate for the lack of labeled samples in low-resource languages (Chinese and Vietnamese). Specifically, we distill the event feature knowledge obtained from the hypergraph channel and the relationship feature knowledge obtained from the line graph channel into the corresponding hypergraph channel and line graph channel in the student model. Furthermore, we introduce cross-perspective self-supervised learning during the student model training process to align the event embeddings of the hypergraph channel and the line graph channel.

[0011] The technical solution of the present invention is: a cross-language news core event analysis method based on graph knowledge distillation, the specific steps of the method include:

[0012] Step 1, construct event hypergraph and event line graph;

[0013] Step 2: Build graph convolution. Graph convolution includes three modules: event hypergraph convolution, line graph convolution, and cross-view self-supervised learning. Graph convolution is used to encode event text to extract higher-dimensional event semantic information and inter-event structural information.

[0014] Step 3: Perform cross-language graph knowledge distillation using a cross-language graph knowledge distillation model. The overall framework of the cross-language graph knowledge distillation model includes two core modules: the event feature knowledge distillation module and the relationship feature knowledge distillation module.

[0015] Cross-language graph knowledge distillation is used to transfer knowledge from the teacher model of a rich-resource language to the student model of a low-resource language using event feature knowledge distillation and relation feature knowledge distillation;

[0016] Specifically, the event feature knowledge extracted by the teacher model through the event hypergraph channel and the relationship knowledge captured through the event line graph channel will be distilled into the corresponding event hypergraph channel and line graph channel in the student model respectively; to enhance the representation ability of the student model, a cross-perspective self-supervised learning mechanism is introduced during the training process to align the event embeddings generated by the hypergraph and line graph channels, which is used to enrich the final event representation of low-resource language news texts. Based on this representation, the core nature of events in news texts can be further analyzed.

[0017] Furthermore, the specific steps of Step 1 include:

[0018] Step 1.1, build event hypergraph:

[0019] Step 1.1.1, express the event hypergraph, that is, EHG graph, as G a =(v a ,ε e ), where v a is a node, ε e is a hyperedge, set the event as a hyperedge, and represent each hyperedge as a i =[m i ,s i ,o i ,t i ,l i ]∈ε e , where m i represents the event mention of the i-th event, s i Represents the event subject of the i-th event, o i represents the event object of the i-th event, t i represents the time when the i-th event occurs, l i Indicates the location where the i-th event occurs; each node is represented by a i,j ∈v a , where a i,jis the i-th event parameter in the j-th document;

[0020] Step 1.1.2, then use an association matrix A∈{0,1} |a|×|e| To represent the links between event nodes, a is the event element, e represents any hyperedge in the hypergraph;

[0021] Step 1.2, build an event line diagram:

[0022] Based on the event hypergraph, it is converted into an event line graph, namely ELG graph, which is used to effectively model the interaction between event hyperedges and further model the deep semantic information and structural information within and between events. The construction of ELG graph is based on the following ideas: the hyperedges in the event hypergraph are regarded as nodes in the line graph, and the two event nodes in the line graph are connected if and only if the two corresponding event hyperedges share the same event element. Given an event hypergraph G a =(v a ,ε e ) is represented by a line graph of L(G)=G e =(v L ,ε L ), where v L ={v e |v e ∈ε e} represents the vertex of the line graph, Representing hyperedges of line graphs, the dependencies between events are modeled as line graphs of event hypergraphs;

[0023] Step 1.3: Embed the input text, EHG graph, and ELG graph nodes based on the mBERT multilingual word embedding model, mapping texts in different languages into a unified semantic space. Specifically:

[0024] Get BERT embedding for each token in the input text, then use bidirectional LSTM to encode the token of the event element of each event, generate a fixed-size event element representation, and finally obtain the element embedding m of all events i ,s i ,o i ,t i ,l i and event embedding i =[m i ,s i ,o i ,t i ,l i ]; It should be noted that in the process of training the student model, mBERT is used to map texts in different languages to the same semantic space.

[0025] Furthermore, the Step 2 includes:

[0026] Step 2.1. Use the hypergraph convolutional network to capture the high-order relationships between event elements to obtain event embedding and model the complex relationships between events. Define the hypergraph convolution as:

[0027]

[0028] Where σ(·) is a nonlinear activation function, including LeakyReLU or eLU, It is a hypergraph convolution of layer l+1, It is l A node attribute of the layer, N1 represents the total number of event element nodes in the event hypergraph, M1 represents the total number of event hyperedges in the event hypergraph, is the adjacency matrix of the event; the filter Θ is used to extract the features of the nodes in the hypergraph. In the specific implementation, the weight of each hyperedge is set to 1, that is, The above formula is normalized and expressed as:

[0029]

[0030] in, represents the row normalization matrix of the l+1 layer, W represents the weight matrix, D e and D v The diagonal matrices representing edge degree and vertex degree respectively;

[0031] Hypergraph convolution efficiently captures high-order semantic relationships in event hypergraphs by performing “node-hyperedge-node” feature transformation on the hypergraph; specifically, the multiplication operation The information aggregation from vertex to hyperedge is defined, and then A is aggregated from hyperedge to vertex; the initial event element vertex feature X( 0 ) After L layers of hypergraph convolution, the node features obtained by each layer of hypergraph convolution are averaged to obtain the final event element vertex embedding X a , the specific calculation formula is as follows:

[0032]

[0033] In the event hypergraph, event element nodes of the same event are connected by an event hyperedge. Therefore, the hyperedge information is constructed by aggregating the embeddings of all nodes connected by the hyperedge. After obtaining the event element embeddings, the hyperedge representation is obtained by mean pooling all these embeddings, and then the final event representation is obtained:

[0034]

[0035] in, represents the embedding of event element hyperedges, N is the set of event element nodes connected by hyperedges, and n represents the total number of event elements connected by hyperedges. is the embedding of event feature nodes that is the input to the l-th hypergraph convolutional layer.

[0036] Furthermore, the Step 2 further includes:

[0037] Step 2.2, based on the hypergraph convolution, introduce the line graph neural network to capture the interaction information between event hyperedges, and further model the deep semantic information and structural information of the event hypergraph; introduce the graph attention GAT to further model the correlation features between events; in the specific implementation, use the strategy used in Hyper-SAGNN to improve the embedding of events; specifically, first according to the weight matrix W Q , W K and W V Calculate the attention coefficient of the node:

[0038]

[0039] Among them, h i and h j denote the embedding of line graph nodes i and j respectively, β ij It represents the voting score of line graph node j to line graph node i, that is, the attention coefficient of line graph node j to line graph node i, by only calculating node j∈N i β ij , to introduce mask attention, where N i Represents the neighborhood of node i in the graph. To facilitate calculation, the softmax function is used to normalize the coefficients:

[0040]

[0041] Finally, the weighted sum of the transformed features with the activation function is calculated by the following formula:

[0042]

[0043] Among them, w represents the weight parameter of the model, W V Represents the median weight matrix of the multi-head attention mechanism;

[0044] A multi-head attention mechanism is introduced to further enhance the model's ability to represent and generalize event graphs. The specific calculation formula is as follows:

[0045]

[0046] Where K represents the total number of heads in the multi-head attention mechanism;

[0047] In each convolution step, an event aggregates information from its neighbors, and the initialized event embedding is input into L layers of graph convolution to capture event-level information; the final representation of event embedding is calculated by averaging the embeddings obtained in each layer, expressed as:

[0048]

[0049] Among them, h (l) Represents the event representation of the lth layer in the line graph convolution, and L represents the total number of line graph convolution layers;

[0050] Step 2.3: The event hypergraph and event line graph are convolved through L layers of graph to obtain the event embedding from the perspective of the event hypergraph. and event embedding from the perspective of event line graph And then get the final embedding of the event:

[0051]

[0052] Here, β is a hyperparameter that balances the event representations of the two channels.

[0053] Furthermore, in Step 2, the cross-view self-supervised learning (CSSL) task is introduced to learn the event hypergraph perspective and the event line graph perspective to ultimately achieve consistency in event representation. In the specific implementation, the InfoNCE loss function is used to maximize the mutual information between the two views, thereby improving the model's performance in event representation learning:

[0054]

[0055] where sim(·,·) measures the cosine similarity between two vectors, |e| is the total number of events in the document, e1 is the ordinal number of an event in the document, and τ e A temperature hyperparameter.

[0056] Furthermore, in Step 3, event feature knowledge distillation is first performed, specifically including:

[0057] Use pre-trained cross-language word embeddings to map low-resource languages into the semantic space of rich-resource languages. Specifically, first, use the Softmax function with a temperature coefficient τ to transform the event representations obtained by the hypergraph modeling of the teacher model and the student model. Convert to soft label To obtain more information within and between events, the soft labels of the student model event representation are also obtained For ease of writing, it is simplified to e a ;

[0058]

[0059] in, represents the soft label of the event representation obtained by the teacher model event hypergraph modeling, e ai and e aj They represent the event representations of event nodes i and j in the event hypergraph channel respectively, τ represents the temperature coefficient hyperparameter, which is used to smooth the output of the teacher model event features, and N(i) represents the set of adjacent nodes of node j. The distillation loss of event feature knowledge distillation is expressed as:

[0060]

[0061] Where D(·) represents the distance function, and M represents the total number of events in the news text;

[0062] In addition, in order to further improve the performance of the student model, the real labels are used to further train the student model during model training; specifically, according to the final embedding e of the event (l) , using a Softmax classification layer:

[0063]

[0064] Among them, W1 and b are learned parameters, and the training process of the model involves using binary cross entropy loss L CE :

[0065]

[0066] Among them, y represents the true label of the sample, Y is the true label set, represents the predicted value of the sample, is a set of predicted values; therefore, the event feature knowledge distillation loss is written as a weighted average of the cross entropy loss and the distillation loss:

[0067] L EKD =(1-α)L CE +αL KD ,

[0068] Here, α is a hyperparameter that balances the cross entropy loss and the distillation loss.

[0069] Furthermore, in Step 3, relational feature knowledge distillation is performed, specifically including:

[0070] An innovative knowledge distillation strategy is proposed, which transfers the relational knowledge learned in the teacher model's line graph channel to the corresponding line graph channel of the student model. This strategy aims to strengthen the student model's ability to model the global semantic information of events, thereby more accurately capturing and representing the complex relationships between events. Specifically, an association matrix is designed to encode the relational knowledge between different event nodes from a pre-trained line graph neural network, and a type-dependent attention layer is introduced to automatically learn the importance of different event nodes.

[0071] Based on the line graph neural network, the event representation of the L-th layer teacher model is obtained and event representation of the student model The relational association matrix is designed to encode the relationship knowledge between different event nodes, which is specifically expressed as:

[0072]

[0073] Among them, i, j represent different event nodes, Used to measure the similarity between two event nodes. The larger the distance between two different event nodes, the farther the distance is. Therefore, the relational feature distillation loss represents:

[0074]

[0075] Where D represents the mean square error of the distance measurement, and M represents the total number of events in the news text.

[0076] Furthermore, in Step 3, the overall loss of the cross-language graph knowledge distillation model is expressed as:

[0077]

[0078] Among them, λ1 and λ2 are hyperparameters that control the weight of relational knowledge distillation and cross-view contrastive learning loss, Θ is all the parameters in the model that can participate in training, and is represented by λ 2Θ Regularization, using Adam as the optimizer during model training.

[0079] The beneficial effects of the present invention are:

[0080] 1. This paper provides a strategy for constructing event hypergraphs and event line graphs. It also distills knowledge from rich-resource models into low-resource models through event feature knowledge distillation and relationship feature knowledge distillation, effectively alleviating the challenges of dealing with low-resource languages. This technology effectively improves the efficiency of analyzing core news events.

[0081] 2. The present invention conducts experiments on three real-world datasets; compared with traditional baseline models, the evaluation indicators of the experimental group are significantly improved compared with all baseline models. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 Constructing schematic diagrams for the event hypergraph and event line graph of the present invention;

[0083] Figure 2 This is the architecture diagram of the cross-language graph knowledge distillation model of the present invention;

[0084] Figure 3 The impact of different model components on model performance on Chinese datasets in this invention;

[0085] Figure 4 The impact of different model components on model performance on Vietnamese datasets of the present invention;

[0086] Figure 5 The impact of the number of convolutional layers on the model performance on the Chinese dataset of the present invention;

[0087] Figure 6 The impact of the number of convolutional layers on the model performance on the Vietnamese dataset of the present invention;

[0088] Figure 7 The impact of hyperparameters on model performance on Chinese datasets in this invention;

[0089] Figure 8 This is the impact of hyperparameters on model performance on the Vietnamese dataset of the present invention. DETAILED DESCRIPTION

[0090] Example 1: Cross-language news core event analysis method based on graph knowledge distillation, such as Figure 1 , which is a schematic diagram of constructing an overall event hypergraph and event line graph of a cross-language news core event analysis method based on graph knowledge distillation provided by an embodiment of the present invention. Figure 2 This is a diagram of the knowledge distillation model architecture of a cross-language news core event analysis method based on graph knowledge distillation. The method includes the following steps:

[0091] Step 1: Construct an event hypergraph and event line graph. By analyzing the characteristics of events in news texts, we found that events in news texts have complex relationships, such as dispersed event elements (event elements are distributed in different sentences or paragraphs) and multi-event element overlap (multiple events share the same event element). Event hypergraphs and event line graphs can effectively model complex interactions within and between events, thereby capturing local and global semantic information about events.

[0092] Furthermore, the specific steps of Step 1 include:

[0093] Step 1.1, build event hypergraph:

[0094] Step 1.1.1, express the event hypergraph, that is, EHG graph, as G a =(ν a ,ε e ), where ν a is a node, ε e is a hyperedge, set the event as a hyperedge, and represent each hyperedge as a i =[m i ,s i ,o i ,t i ,l i ]∈ε e , where m i represents the event mention of the i-th event, s i Represents the event subject of the i-th event, o i represents the event object of the i-th event, t i represents the time when the i-th event occurs, l i Indicates the location where the i-th event occurs; each node is represented by a i,j ∈ν a , where a i,j is the i-th event parameter in the j-th document;

[0095] Step 1.1.2, then use an association matrix A∈{0,1} |a|×|e| To represent the links between event nodes, a is the event element, e represents any hyperedge in the hypergraph;

[0096] like Figure 1 The event hypergraph described in the above example, event e1 contains event elements a 1,1 ,a 1,2 ,a 1,3 ,a 1,4 , events e1 and e2 share event element a 1,2 It is worth emphasizing that the event hypergraph mainly captures the high-order semantic relationships between event elements, thereby modeling the global semantic information of events. To a certain extent, it also solves the difficulty of semantic modeling long texts in traditional text encoders.

[0097] Step 1.2, build an event line diagram:

[0098] Hypergraphs can model high-order semantic relationships and long-distance dependencies between event elements. However, there are often complex associations between events in news texts, such as co-reference, causality, temporal sequence, and parent-child relationships. Modeling the association features between the above events is crucial to measuring whether an event is a key event in the news text. In addition, studies have shown that line graphs can effectively model the implicit structural information between nodes in the graph. Therefore, based on the event hypergraph, it is converted into an event line graph, namely an ELG graph, which is used to effectively model the interaction between event hyperedges and further model the deep semantic information and structural information within and between events; the construction of the ELG graph is based on the following idea: the hyperedges in the event hypergraph are regarded as nodes in the line graph, and the two event nodes in the line graph will be connected if and only if the two corresponding event hyperedges share the same event element; given an event hypergraph G a =(v a ,ε e ) is represented by a line graph of L(G)=G e =(v L ,ε L ), where ν L ={v e |v e ∈ε e} represents the vertex of the line graph, Representing hyperedges of line graphs, the dependencies between events are modeled as line graphs of event hypergraphs;

[0099] Step 1.3: Cross-language core analysis and detection. The first challenge to be solved is how to overcome the differences between different languages, eliminate the language gap between different languages, and achieve the representation and calculation of multilingual news text in a unified space. Therefore, based on the mBERT multilingual word embedding model, the input text, EHG graph, and ELG graph nodes are embedded, mapping texts in different languages into a unified semantic space. Specifically:

[0100] Get BERT embedding for each token in the input text, then use bidirectional LSTM to encode the token of the event element of each event, generate a fixed-size event element representation, and finally obtain the element embedding m of all events i ,s i ,o i ,t i ,l i and event embedding i ={m i ,s i ,o i ,t i ,l i ]; It should be noted that in the process of training the student model, mBERT is used to map texts in different languages to the same semantic space.

[0101] Step 2: Build graph convolution. Graph convolution includes three modules: event hypergraph convolution, line graph convolution, and cross-view self-supervised learning. Graph convolution is used to encode event text to extract higher-dimensional event semantic information and inter-event structural information.

[0102] Furthermore, the Step 2 includes:

[0103] Step 2.1. Use hypergraph convolutional networks to capture high-order relationships between event elements to obtain event embeddings and model the complex relationships between events. The main challenge in defining hypergraph convolution is how to propagate node embeddings to learn hyperedge embeddings. Inspired by the spectral hypergraph convolution proposed by Sandhaus et al., hypergraph convolution is defined as:

[0104]

[0105] Where σ(·) is a nonlinear activation function, including LeakyReLU or eLU, It is a hypergraph convolution of layer l+1, It is l A node attribute of the layer, N1 represents the total number of event element nodes in the event hypergraph, M1 represents the total number of event hyperedges in the event hypergraph, is the adjacency matrix of the event; the filter Θ is used to extract the features of the nodes in the hypergraph. In the specific implementation, the weight of each hyperedge is set to 1, that is, The above formula is normalized and expressed as:

[0106]

[0107] in, represents the row normalization matrix of the l+1 layer, W represents the weight matrix, D e and D v The diagonal matrices representing edge degree and vertex degree respectively;

[0108] Hypergraph convolution efficiently captures high-order semantic relationships in event hypergraphs by performing “node-hyperedge-node” feature transformation on the hypergraph; specifically, the multiplication operation The information aggregation from vertex to hyperedge is defined, and then A is aggregated from hyperedge to vertex; the initial event element vertex feature X( 0 ) After L layers of hypergraph convolution, the node features obtained by each layer of hypergraph convolution are averaged to obtain the final event element vertex embedding X a , the specific calculation formula is as follows:

[0109]

[0110] In the event hypergraph, event element nodes of the same event are connected by an event hyperedge. Therefore, the hyperedge information is constructed by aggregating the embeddings of all nodes connected by the hyperedge. After obtaining the event element embeddings, the hyperedge representation is obtained by mean pooling all these embeddings, and then the final event representation is obtained:

[0111]

[0112] in, represents the embedding of event element hyperedges, N is the set of event element nodes connected by hyperedges, and n represents the total number of event elements connected by hyperedges. is the embedding of event feature nodes that is the input to the l-th hypergraph convolutional layer.

[0113] Furthermore, the Step 2 further includes:

[0114] Step 2.2, event hypergraph modeling enables the model to effectively capture the high-level semantic features between event elements. However, the events described in news texts have strong contextual characteristics, and each specific event establishes a discourse relationship with other events in the news context. For example, Figure 1 In the figure, events e1, e2, and e3 described in the same news text are connected by sharing the same event elements. It is crucial for the model to capture the correlation features between events in order to model the core nature of news events. Therefore, based on the hypergraph convolution, a line graph neural network is introduced to capture the interactive information between event hyperedges, and further model the deep semantic information and structural information of the event hypergraph; another key to modeling the core nature of news events is to capture the interactive voting between events, and events that receive higher votes from other events will be given higher attention scores. To this end, the graph attention network (GAT) is introduced to further model the correlation features between events; in the specific implementation, the strategy used in Hyper-SAGNN is used to improve the embedding of events; specifically, first, according to the weight matrix W Q , W K and W V Calculate the attention coefficient of the node:

[0115]

[0116] Among them, h i and h j denote the embedding of line graph nodes i and j respectively, β ij It represents the voting score of line graph node j to line graph node i, that is, the attention coefficient of line graph node j to line graph node i, by only calculating node j∈N i β ij , to introduce mask attention, where Ni Represents the neighborhood of node i in the graph. To facilitate calculation, the softmax function is used to normalize the coefficients:

[0117]

[0118] Finally, the weighted sum of the transformed features with the activation function is calculated by the following formula:

[0119]

[0120] Among them, w represents the weight parameter of the model, W V Represents the median weight matrix of the multi-head attention mechanism;

[0121] A multi-head attention mechanism is introduced to further enhance the model's ability to represent and generalize event graphs. The specific calculation formula is as follows:

[0122]

[0123] Where K represents the total number of heads in the multi-head attention mechanism;

[0124] In each convolution step, an event aggregates information from its neighbors, and the initialized event embedding is input into L layers of graph convolution to capture event-level information; the final representation of event embedding is calculated by averaging the embeddings obtained in each layer, expressed as:

[0125]

[0126] Among them, h (l) Represents the event representation of the lth layer in the line graph convolution, and L represents the total number of line graph convolution layers;

[0127] Step 2.3: The event hypergraph and event line graph are convolved through L layers of graph to obtain the event embedding from the perspective of the event hypergraph. and event embedding from the perspective of event line graph And then get the final embedding of the event:

[0128]

[0129] Here, β is a hyperparameter that balances the event representations of the two channels.

[0130] Furthermore, in Step 2, although we represent the final embedding of the event as a synthesis of two views, the consistency between the two views is omitted. In theory, to improve the generalization and robustness of the model, we should guide model learning to ensure that the embedding representations of the same event in different hypergraph convolution channels are as close as possible, and vice versa. To address this limitation, the cross-view self-supervised learning (CSSL) task is introduced to learn the event hypergraph perspective and the event line graph perspective to ultimately achieve consistency in event representation; in the specific implementation, the InfoNCE loss function is used to maximize the mutual information between the two views, thereby improving the model's performance in event representation learning:

[0131]

[0132] where sim(·,·) measures the cosine similarity between two vectors, |e| is the total number of events in the document, e1 is the ordinal number of an event in the document, and τ e A temperature hyperparameter.

[0133] Step 3: Perform cross-language graph knowledge distillation using a cross-language graph knowledge distillation model. The overall framework of the cross-language graph knowledge distillation model includes two core modules: the event feature knowledge distillation module and the relationship feature knowledge distillation module.

[0134] Cross-language graph knowledge distillation is used to transfer knowledge from the teacher model of a rich-resource language to the student model of a low-resource language using event feature knowledge distillation and relation feature knowledge distillation;

[0135] Specifically, the event feature knowledge extracted by the teacher model through the event hypergraph channel and the relationship knowledge captured through the event line graph channel will be distilled into the corresponding event hypergraph channel and line graph channel in the student model respectively; to enhance the representation ability of the student model, a cross-perspective self-supervised learning mechanism is introduced during the training process to align the event embeddings generated by the hypergraph and line graph channels, which is used to enrich the final event representation of low-resource language news texts. Based on this representation, the core nature of events in news texts can be further analyzed.

[0136] Furthermore, in Step 3, event feature knowledge distillation is first performed, specifically including:

[0137] Distilling the event feature knowledge learned from the hypergraph modeling of rich-resource languages to low-resource languages can effectively improve the performance of the news core analysis and detection model in low-resource languages. Since the teacher model and the student model belong to different language modalities, the cross-language problem needs to be solved first. Therefore, pre-trained cross-language word embeddings are used to map the low-resource language to the semantic space of the rich-resource language. Specifically, the event representations obtained by the teacher model and the student model hypergraph modeling are firstly transformed into the semantic space of the rich-resource language through the Softmax function with a temperature coefficient τ. Convert to soft label To obtain more information within and between events, the soft labels of the student model event representation are also obtained For ease of writing, it is simplified to e a ;

[0138]

[0139] in, represents the soft label of the event representation obtained by the teacher model event hypergraph modeling, e ai and e aj They represent the event representations of event nodes i and j in the event hypergraph channel respectively, τ represents the temperature coefficient hyperparameter, which is used to smooth the output of the teacher model event features, and N(i) represents the set of adjacent nodes of node j. The distillation loss of event feature knowledge distillation is expressed as:

[0140]

[0141] Where D(·) represents the distance function, and M represents the total number of events in the news text;

[0142] In addition, in order to further improve the performance of the student model, the real labels are used to further train the student model during model training; specifically, according to the final embedding e of the event (l) , using a Softmax classification layer:

[0143]

[0144] Among them, W1 and b are learned parameters, and the training process of the model involves using binary cross entropy loss L CE :

[0145]

[0146] Among them, y represents the true label of the sample, Y is the true label set, represents the predicted value of the sample, is a set of predicted values; therefore, the event feature knowledge distillation loss is written as a weighted average of the cross entropy loss and the distillation loss:

[0147] L EKD =(1-α)L CE +αL KD ,

[0148] Here, α is a hyperparameter that balances the cross entropy loss and the distillation loss.

[0149] Furthermore, in Step 3, relational feature knowledge distillation is performed, specifically including:

[0150] In the process of knowledge distillation, only focusing on the knowledge at the event level is transferred to the student model, which makes it difficult to effectively capture the complex correlation features between events, which are crucial for accurately modeling the saliency of events. Through in-depth analysis, we observed that in the teacher model's line graph channel, the intermediate convolutional layer aggregates the correlation relationships between the embeddings of different event nodes. These correlation relationships play a key role in modeling the saliency of events. They directly reveal how the teacher model captures and models high-order relationships between different nodes in the embedding feature space. Based on this, an innovative knowledge distillation strategy is proposed, which transfers the relational knowledge learned in the teacher model's line graph channel to the corresponding line graph channel of the student model, aiming to enhance the student model's ability to model the global semantic information of events, thereby more accurately capturing and representing the complex relationships between events. Specifically, an association matrix is designed to encode the relational knowledge between different event nodes from the pre-trained line graph neural network, and a type-related attention layer is introduced to automatically learn the importance of different event nodes.

[0151] Based on the line graph neural network, the event representation of the L-th layer teacher model is obtained and event representation of the student model The relational association matrix is designed to encode the relationship knowledge between different event nodes, which is specifically expressed as:

[0152]

[0153] Among them, i, j represent different event nodes, Used to measure the similarity between two event nodes. The larger the distance between two different event nodes, the farther the distance is. Therefore, the relational feature distillation loss represents:

[0154]

[0155] Where D represents the mean square error of the distance measurement, and M represents the total number of events in the news text.

[0156] Furthermore, in Step 3, the overall loss of the cross-language graph knowledge distillation model is expressed as:

[0157]

[0158] Among them, λ1 and λ2 are hyperparameters that control the loss weights of relational knowledge distillation and cross-view contrastive learning, and θ is all the parameters in the model that can participate in training, which is determined by λ 2Θ Regularization, using Adam as the optimizer during model training.

[0159] To illustrate the effectiveness of the present invention, we compared it with existing methods and conducted tests on three datasets. The evaluation indicators used were precision P@k, recall R@k, and normalized recall NR@k, where k = 1, 5, and 10, and the calculation formula is as follows:

[0160]

[0161] Experiment 1: Comparative experiment

[0162] Table 1 Experimental results of English, Chinese and Vietnamese on pre-trained teacher models

[0163]

[0164]

[0165] From the experimental results presented in Table 1, we can clearly observe that the current state-of-the-art models KCE and CEE-IEA achieve remarkable performance on the news core analysis detection task on the English dataset, with AUC values reaching 69.76% and 73.49%, respectively. However, when applied to the Chinese and Vietnamese datasets, despite varying degrees of performance degradation across various evaluation metrics, the baseline model CEE-IEA still achieves relatively impressive results on the AUC evaluation metric, achieving 64.34% (a decrease of 9.15% compared to the English dataset) and 59.53% (a decrease of 13.96% compared to the English dataset), respectively. Furthermore, compared to the corresponding experimental results in Tables 2 and 3, the experimental results after fine-tuning the student model outperform those in Table 1. These experimental results strongly demonstrate the effectiveness of pre-trained teacher models in knowledge transfer, providing rich prior knowledge to the student model, thereby providing possibilities and directions for improving the performance of news core analysis detection models in low-resource language environments.

[0166] Compared to the baseline models KCE and CEE-IEA, the proposed model achieved the best results across all languages. Specifically, on the key evaluation metric AUC, the proposed model achieved a 6.87% improvement in English, a 6.52% improvement in Chinese, and a 5.24% improvement in Vietnamese. This series of experimental data not only demonstrates the effectiveness of the proposed model but also further validates the effectiveness of hypergraph and line graph structures in modeling global semantic information about events and inter-event correlation features.

[0167] Table 2 Performance comparison results of different baseline models on Chinese datasets

[0168]

[0169]

[0170] Table 3 Performance comparison results of different baseline models on Vietnamese dataset

[0171] Model P@1 P@5 P@10 R@1 R@5 R@10 NR@1 NR@5 NR@10 AUC Location 43.18 35.59 26.98 8.51 30.89 45.21 44.18 52.20 61.84 55.03 PageRank 45.51 36.02 28.27 8.59 32.84 47.59 47.00 52.10 61.87 62.01 Frequency 50.08 36.81 27.19 9.28 34.52 52.27 51.27 51.68 60.59 62.04 LPF 51.97 37.80 28.65 9.89 35.69 52.82 52.89 52.92 62.81 63.17 KCE 58.87 40.68 29.02 8.57 36.82 54.82 58.25 59.51 68.21 66.83 CEE-IEA 60.81 42.97 31.29 10.08 39.82 56.73 61.70 62.08 71.62 70.25 R-GCN 60.67 42.08 30.85 12.08 39.72 57.21 60.82 62.32 72.93 71.50 GCN-SED 62.82 44.94 35.09 14.08 41.87 60.59 61.92 63.82 73.29 72.07 GAT-SED 65.02 46.80 35.93 14.87 42.19 60.83 63.08 64.29 73.82 72.53 GKDT(Ours) 67.69 49.01 36.89 16.65 44.82 63.33 68.14 70.08 78.45 <![CDATA[79.51 * ]]> Δ 2.67 2.21 0.96 1.78 2.63 2.5 5.06 5.79 4.63 6.98

[0172] From the experimental results in Table 2 and Table 3, we can see that:

[0173] (1) In the evaluation of a baseline model based entirely on features (including Location, PageRank, and Frequency), we observed that Frequency and PageRank outperformed Location. This finding reveals the central role that event trigger words and event elements in news core events play in news reports, running through the entire report from beginning to end. Furthermore, we infer that the importance of an event can be effectively modeled by the frequency of occurrence of event elements in the document and their correlation with the event.

[0174] (2) In a comprehensive comparison, neural network-based methods significantly outperformed completely feature-based methods in overall performance. In particular, the KCE method demonstrated excellent performance across all evaluation metrics, particularly in the P@1 and NR@1 metrics, where performance improved by up to 10 percentage points, significantly surpassing benchmark models such as Location, PageRank, and Frequency. Furthermore, the experimental results demonstrate the importance of modeling inter-event correlation features for improving model performance.

[0175] (3) In the comparative experiments, graph neural network-based methods (such as R-GCN, GCN-SED, and GAT-SED) achieved better results than other baseline models. This achievement is mainly attributed to their ability to efficiently aggregate event attribute features and structural information. Among them, the performance of GAT-SED is particularly outstanding, which shows the superiority of graph attention networks in calculating attention weights to further model event importance features. It is worth noting that the performance of the R-GCN model in Vietnamese is slightly reduced. After analysis, we speculate that this may be because R-GCN constructs a news document association graph based on syntactic structure, while the performance of Vietnamese syntactic parsing function needs to be improved, resulting in error cascades.

[0176] (4) In the cross-language scenario, the GKDT proposed in the present invention achieved the best experimental results among all baseline models. Specifically, in the Chinese and Vietnamese environments, the AUC evaluation indicators achieved an improvement of 6.98% and 3.08%, respectively. Further compared with the experimental results in Table 1, the average performance improvement of the GKDT model in Chinese and Vietnamese was as high as 8.69% and 10.17%, respectively. These remarkable results verify the effectiveness of the event feature knowledge distillation and relation feature knowledge distillation strategies, which successfully transferred the prior knowledge in the teacher model to the student model, thereby significantly improving the performance of the low-resource language model. It is worth mentioning that compared with the performance in the Chinese scenario, the performance of GKDT in Vietnamese is even better, which fully demonstrates that the method proposed in the present invention is highly competitive in low-resource language scenarios.

[0177] Experiment 2: Feature Ablation Experiment

[0178] To ensure a fair and accurate comparison, following the research paradigm of Jindal et al., we incorporated global event features into our model implementation. These global features include key elements such as Frame Name, Sentence Location, Trigger Frequency, Argument Frequency, Named Argument, and Parent Score. To further explore the actual role and impact of these global event features in the model, we will conduct an ablation study on these global event features. Specifically, we will evaluate the impact of these features on model performance by comparing GKDT with its simplified variants. The experimental results are shown in Table 4, where ▽ represents the impact of different features on the performance of the GKDT model using the evaluation metric AUC. Bold numbers represent the best experimental results.

[0179] (1) w / o Parent Score (PS): indicates the number of sub-events that are not incorporated into a specific event in the model.

[0180] (2) w / o Frame Name (FN): indicates that the event frame is not integrated into the model, that is, the abstract description of the event is not integrated.

[0181] (3) w / o Sentence Location (SL): means that the sentence location feature is not incorporated into the model.

[0182] (4) w / o Event Trigger Frequency (TF): means that events are not incorporated into the model

[0183] (5) w / o Argument Frequency (AF): indicates the frequency of occurrence of event elements that are not incorporated into the model.

[0184] (6) w / o Named Argument (NA): indicates that the entity characteristics of the event element are not incorporated into the model.

[0185] Table 4 Impact of different event global features on model performance

[0186]

[0187]

[0188] The experimental results in Table 4 show that: (1) For both Chinese and Vietnamese, when the global event features are incorporated into the model, the performance of the model is significantly improved. Specifically, in terms of AUC, an important indicator for evaluating model performance, the Chinese and Vietnamese models achieved gains of 0.91% and 1.15%, respectively. This finding shows that global features are crucial for improving the effectiveness of cross-language event analysis tasks. (2) In another evaluation indicator, NR@k, as the k value increases, we find that the role of TF (event trigger word frequency) and AF (event element frequency) features in the global event features is gradually increasing. This observation further reveals the important value of TF and AF features in distinguishing salient events from other events, which helps the model to more accurately identify and distinguish different events in complex event analysis tasks.

[0189] Experiment 3: Model Component Ablation Experiment

[0190] In the ablation experiment, we mainly analyze the impact of different features and different module implementation structures on the performance of cross-language news core analysis and detection.

[0191] This paper mainly explores the impact of different component implementations in the GKDT model on the cross-language news core analysis and detection performance. The proposed GKDT model is compared with the following variant models:

[0192] (1) GKDT(-CL): indicates that the contrastive learning module is not used in the model to learn the consistency of event representation of hypergraph channel and line graph channel.

[0193] (2) GKDT(-ATT): indicates that GAT is not used in the line graph convolution in the model to model the association features between event nodes.

[0194] (3) GKDT(-EKD): It means removing the event feature knowledge distillation component in the model knowledge distillation stage and only retaining the cross entropy loss in the student model, that is, L EKD =L CE .

[0195] (4)GKDT(-RKD): indicates the removal of relational feature knowledge distillation components in the model knowledge distillation stage.

[0196] Depend on Figure 3 and Figure 4 The experimental results in , we can conclude that:

[0197] (1) In experiments with Chinese and Vietnamese corpora, the proposed GKDT model significantly outperformed all variant models on key evaluation metrics (including NR@1, NR@5, NR@10, and AUC). Furthermore, when each component of the model was removed one by one, the model performance showed a downward trend, which fully verified the effectiveness and indispensability of each component in improving the overall performance.

[0198] (2) During the student model training process, we observed a decrease in model performance when contrastive learning and graph attention mechanisms were removed. This finding suggests that: first, contrastive learning plays a key role in capturing deep event semantic information, effectively alleviating the problems caused by data imbalance and sparsity of event hypergraphs; second, graph attention mechanisms can effectively model the complex relationships between events, which has a significant impact on the task of analyzing core news events.

[0199] (3) During student model training, removing both event feature knowledge distillation (EKD) and relational knowledge distillation (RKD) significantly impacts model performance. This highlights the importance of modeling both EKD and RKD simultaneously. By distilling the event feature knowledge and inter-event relation knowledge learned from the teacher model into the student model, the performance of the student model can be significantly improved.

[0200] (4) Comparing the two knowledge distillation methods (EKD and RKD), we found that EKD has a relatively limited improvement on model performance, while RKD shows a more significant performance improvement, especially in the extremely low-resource language environment such as Vietnamese. The above experimental results verify the transferability of event correlation features and emphasize their important role in improving the performance of student models in low-resource language scenarios.

[0201] Experiment 4: Analysis of the impact of the number of convolutional layers on model performance

[0202] In order to deeply explore the specific impact of different numbers of convolutional layers on model performance, this paper presents in detail the experimental results of the GKDT model on Chinese and Vietnamese datasets when using different numbers of convolutional layers. Figure 5 and Figure 6 As shown in the figure, on the Chinese dataset, when the number of convolutional layers of the GKDT model is set to 2, it significantly outperforms other configurations in all evaluation indicators, showing the best performance. In addition, the GKDT model shows relatively low sensitivity to changes in the number of convolutional layers on the Chinese dataset. However, on the Vietnamese dataset, the model performance shows a completely different trend. Specifically, when the number of convolutional layers is set to 3, the GKDT model shows the best performance, and the impact of different numbers of convolutional layers on model performance is more significant. Furthermore, with the increase in the number of convolutional layers, the performance of the GKDT model on various evaluation indicators shows a significant downward trend. One potential explanation for this phenomenon is that as the depth of the model increases, the GKDT model may be more susceptible to the over-smoothing problem when aggregating high-order information from distant neighbors, resulting in a gradual decline in its performance.

[0203] Experiment 5: Model Hyperparameter Analysis

[0204] Several hyperparameters are tuned on the Chinese and Vietnamese datasets to verify the sensitivity of the GKDT model. More specifically, we use grid search to select the formula The best hyperparameters τ, α, λ1, λ2 in the experiment are shown in the following table. Figure 7 and Figure 8 shown.

[0205] Depend on Figure 7 and Figure 8 The experimental results in , we can conclude that:

[0206] (1) Softmax temperature coefficient τ: According to the formula It can be seen that the higher the temperature coefficient τ value, the smoother the probability distribution of event node categories. Figure 7 (a) and Figure 8From the experimental results of (a), it can be observed that with the gradual increase of the temperature coefficient τ value, the performance of the GKDT model corresponding to the method of the present invention also shows a steadily increasing trend. In particular, when the τ value is set to 8, the model achieves optimal performance in multiple evaluation indicators such as NR@1, NR@5, NR@10 and AUC, which fully demonstrates the positive impact of appropriately increasing the temperature coefficient on model performance. However, when the temperature coefficient τ value exceeds 8, the model performance begins to show a slow downward trend. After analysis, we believe that this may be due to the excessively high temperature coefficient causing the model to be overly smooth, thereby weakening its ability to distinguish different event node categories to a certain extent.

[0207] (2) Weight hyperparameter α: Formula The weight hyperparameter in is used to balance the importance of cross entropy loss and event feature knowledge distillation loss. Figure 7 (b) and Figure 8 As can be observed from the experimental results in (b), we found that the GKDT model is not particularly sensitive to changes in the hyperparameter α in general. However, it is worth noting that on the Vietnamese dataset, the model's sensitivity to α is enhanced compared to the Chinese dataset. Specifically, on the Chinese dataset, when α is set to 0.6, the model performance reaches its peak; on the Vietnamese dataset, when the α value is 0.8, the model performance reaches the best, and as the value increases, the model performance fluctuates greatly. This phenomenon not only verifies the important role of α in adjusting the model loss function, but also further confirms the effectiveness and applicability of the event feature knowledge distillation method proposed in this invention in low-resource language scenarios.

[0208] (3) Weight hyperparameter λ1: In order to further explore the specific impact of relational knowledge distillation on model performance, we systematically analyzed The performance of the model under different values of the weight coefficient λ1 is as follows: Figure 7 (c) and Figure 8 (c) As shown. As the value of λ1 increases, the performance of the GKDT model shows an upward trend. Specifically, on the Chinese and Vietnamese datasets, when the λ1 value is set to 100 and 110, respectively, the GKDT model achieves optimal performance. Further analysis shows that on the Chinese dataset, when the λ1 value is in the range of [0.01, 1], the GKDT model exhibits relatively stable performance; on the Vietnamese dataset, the model performance continues to improve with the increase of the λ1 value. These experimental results fully demonstrate the strategy of effectively distilling the relational knowledge from the resource-rich language teacher model into the low-resource language student model, significantly reducing the dependence of the news core event analysis model on labeled data.

[0209] (4) Weight hyperparameter λ2: In the formula A weight hyperparameter λ2 is introduced in to control the magnitude of cross-view self-supervised learning. To further explore the impact of cross-view self-supervised learning on model performance, we designed a set of experiments in which the value range of λ2 is set to [0.00, 0.01, 0.02, 0.03, 0.05] and Figure 7 (d) and Figure 8 (d) shows the model performance under different values. Analysis of the experimental results shows that on the Chinese dataset, the GKDT model achieves optimal performance when λ2 is set to 0.02; on the Vietnamese dataset, the optimal value of λ2 is 0.03. It is worth noting that by comparing the results on the two language datasets, we found that cross-perspective self-supervised learning (CSSL) significantly improves model performance on the Vietnamese dataset. This further verifies that the CSSL strategy proposed in this paper has significant advantages in low-resource language scenarios and can effectively reduce the model's dependence on large amounts of labeled data.

[0210] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A cross-language news core event analysis method based on graph knowledge distillation, characterized by: The specific steps of the method include: Step 1, construct event hypergraph and event line graph; Step 2: Build graph convolution. Graph convolution includes three modules: event hypergraph convolution, line graph convolution, and cross-view self-supervised learning. Graph convolution is used to encode event text to extract higher-dimensional event semantic information and inter-event structural information. Step 3: Perform cross-language graph knowledge distillation using a cross-language graph knowledge distillation model. The overall framework of the cross-language graph knowledge distillation model includes two core modules: the event feature knowledge distillation module and the relationship feature knowledge distillation module. Cross-language graph knowledge distillation is used to transfer knowledge from the teacher model of a rich-resource language to the student model of a low-resource language using event feature knowledge distillation and relation feature knowledge distillation; Specifically, the event feature knowledge extracted by the teacher model through the event hypergraph channel and the relationship knowledge captured through the event line graph channel will be distilled into the corresponding event hypergraph channel and line graph channel in the student model respectively; to enhance the representation ability of the student model, a cross-perspective self-supervised learning mechanism is introduced during the training process to align the event embeddings generated by the hypergraph and line graph channels, which is used to enrich the final event representation of low-resource language news texts. Based on this representation, the core nature of events in news texts can be further analyzed.

2. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1 is characterized by: The specific steps of Step 1 include: Step 1.1, build event hypergraph: Step 1.1.1, express the event hypergraph, that is, EHG graph, as G a =(v a ,ε e ), where v a is a node, ε e is a hyperedge, set the event as a hyperedge, and represent each hyperedge as a i =[m i ,s i ,o i ,t i ,l i ]∈ε e , where m i represents the event mention of the i-th event, s i Represents the event subject of the i-th event, o i represents the event object of the i-th event, t i represents the time when the i-th event occurs, l i Indicates the location where the i-th event occurs; each node is represented by a i,j ∈v a , where a i,j is the i-th event parameter in the j-th document; Step 1.1.2, then use an association matrix A∈{0,1} |a|×|e| To represent the links between event nodes, a is the event element, e represents any hyperedge in the hypergraph; Step 1.2, build an event line diagram: Based on the event hypergraph, it is converted into an event line graph, namely ELG graph, which is used to effectively model the interaction between event hyperedges and further model the deep semantic information and structural information within and between events. The construction of ELG graph is based on the following ideas: the hyperedges in the event hypergraph are regarded as nodes in the line graph, and the two event nodes in the line graph are connected if and only if the two corresponding event hyperedges share the same event element. Given an event hypergraph G a =(v a ,ε e ) is represented by a line graph of L(G)=G e =(v L ,ε L ), where v L ={v e |v e ∈ε e } represents the vertex of the line graph, Representing hyperedges of line graphs, the dependencies between events are modeled as line graphs of event hypergraphs; Step 1.3: Embed the input text, EHG graph, and ELG graph nodes based on the mBERT multilingual word embedding model, mapping texts in different languages into a unified semantic space. Specifically: Get BERT embedding for each token in the input text, then use bidirectional LSTM to encode the token of the event element of each event, generate a fixed-size event element representation, and finally obtain the element embedding m of all events i ,s i ,o i ,t i ,l i and event embedding i =[m i ,s i ,o i ,t i ,l i ]; It should be noted that in the process of training the student model, mBERT is used to map texts in different languages to the same semantic space.

3. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1 is characterized by: Step 2 includes: Step 2.

1. Use the hypergraph convolutional network to capture the high-order relationships between event elements to obtain event embedding and model the complex relationships between events. Define the hypergraph convolution as: Where σ(·) is a nonlinear activation function, including LeakyReLU or eLU, It is a hypergraph convolution of layer l+1, is a node attribute of the lth layer, N1 represents the total number of event element nodes in the event hypergraph, M1 represents the total number of event hyperedges in the event hypergraph, is the adjacency matrix of the event; the filter Θ is used to extract the features of the nodes in the hypergraph. In the specific implementation, the weight of each hyperedge is set to 1, that is, The above formula is normalized and expressed as: in, represents the row normalization matrix of the l+1 layer, W represents the weight matrix, D e and D v The diagonal matrices representing edge degree and vertex degree respectively; Hypergraph convolution efficiently captures high-order semantic relationships in event hypergraphs by performing "node-hyperedge-node" feature transformation on the hypergraph; specifically, the multiplication operation The information aggregation from vertex to hyperedge is defined, and then A is aggregated from hyperedge to vertex; the initial event element vertex feature X( 0 ) After L layers of hypergraph convolution, the node features obtained by each layer of hypergraph convolution are averaged to obtain the final event element vertex embedding X a , the specific calculation formula is as follows: In the event hypergraph, event element nodes of the same event are connected by an event hyperedge. Therefore, the hyperedge information is constructed by aggregating the embeddings of all nodes connected by the hyperedge. After obtaining the event element embeddings, the hyperedge representation is obtained by mean pooling all these embeddings, and then the final event representation is obtained: in, represents the embedding of event element hyperedges, N is the set of event element nodes connected by hyperedges, and n represents the total number of event elements connected by hyperedges. is the embedding of event feature nodes that is the input to the l-th hypergraph convolutional layer.

4. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1 is characterized by: The Step 2 further includes: Step 2.2, based on the hypergraph convolution, introduce the line graph neural network to capture the interaction information between event hyperedges, and further model the deep semantic information and structural information of the event hypergraph; introduce the graph attention GAT to further model the correlation features between events; in the specific implementation, use the strategy used in Hyper-SAGNN to improve the embedding of events; specifically, first according to the weight matrix W Q , W K and W V Calculate the attention coefficient of the node: Among them, h i and h j denote the embedding of line graph nodes i and j respectively, β ij It represents the voting score of line graph node j to line graph node i, that is, the attention coefficient of line graph node j to line graph node i, by only calculating node j∈N i β ij , to introduce mask attention, where N i Represents the neighborhood of node i in the graph. To facilitate calculation, the softmax function is used to normalize the coefficients: Finally, the weighted sum of the transformed features with the activation function is calculated by the following formula: Among them, w represents the weight parameter of the model, W V Represents the median weight matrix of the multi-head attention mechanism; A multi-head attention mechanism is introduced to further enhance the model's ability to represent and generalize event graphs. The specific calculation formula is as follows: Where K represents the total number of heads in the multi-head attention mechanism; In each convolution step, an event aggregates information from its neighbors, and the initialized event embedding is input into L layers of graph convolution to capture event-level information; the final representation of event embedding is calculated by averaging the embeddings obtained in each layer, expressed as: Among them, h (l) Represents the event representation of the lth layer in the line graph convolution, and L represents the total number of line graph convolution layers; Step 2.3: The event hypergraph and event line graph are convolved through L layers of graph to obtain the event embedding from the perspective of the event hypergraph. and event embedding from the perspective of event line graph And then get the final embedding of the event: Here, β is a hyperparameter that balances the event representations of the two channels.

5. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1 is characterized by: In Step 2, the cross-view self-supervised learning (CSSL) task is introduced to learn the event hypergraph and event line graph perspectives to ultimately achieve consistency in event representation. In the specific implementation, the InfoNCE loss function is used to maximize the mutual information between the two views, thereby improving the model's performance in event representation learning: where sim(·,·) measures the cosine similarity between two vectors, |e| is the total number of events in the document, e1 is the ordinal number of an event in the document, and τ e A temperature hyperparameter.

6. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1 is characterized by: In Step 3, event feature knowledge distillation is first performed, specifically including: Use pre-trained cross-language word embeddings to map low-resource languages into the semantic space of rich-resource languages. Specifically, first, use the Softmax function with a temperature coefficient τ to transform the event representations obtained by the hypergraph modeling of the teacher model and the student model. Convert to soft label To obtain more information within and between events, the soft labels of the student model event representation are also obtained For ease of writing, it is simplified to e a ; in, represents the soft label of the event representation obtained by the teacher model event hypergraph modeling, e ai and e aj They represent the event representations of event nodes i and j in the event hypergraph channel respectively, τ represents the temperature coefficient hyperparameter, which is used to smooth the output of the teacher model event features, and N(i) represents the set of adjacent nodes of node j. The distillation loss of event feature knowledge distillation is expressed as: Where D(·) represents the distance function, and M represents the total number of events in the news text; In addition, in order to further improve the performance of the student model, the real labels are used to further train the student model during model training; specifically, according to the final embedding e of the event (l) , using a Softmax classification layer: Among them, W1 and b are learned parameters, and the training process of the model involves using binary cross entropy loss L CE : Among them, y represents the true label of the sample, Y is the true label set, represents the predicted value of the sample, is a set of predicted values; therefore, the event feature knowledge distillation loss is written as a weighted average of the cross entropy loss and the distillation loss: L EKD =(1-α)L CE +αL KD , Here, α is a hyperparameter that balances the cross entropy loss and the distillation loss.

7. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1 is characterized by: In Step 3, relational feature knowledge distillation is performed, specifically including: An innovative knowledge distillation strategy is proposed, which transfers the relational knowledge learned in the teacher model's line graph channel to the corresponding line graph channel of the student model. This strategy aims to strengthen the student model's ability to model the global semantic information of events, thereby more accurately capturing and representing the complex relationships between events. Specifically, an association matrix is designed to encode the relational knowledge between different event nodes from a pre-trained line graph neural network, and a type-dependent attention layer is introduced to automatically learn the importance of different event nodes. Based on the line graph neural network, the event representation of the L-th layer teacher model is obtained and event representation of the student model The relational association matrix is designed to encode the relationship knowledge between different event nodes, which is specifically expressed as: Among them, i, j represent different event nodes, Used to measure the similarity between two event nodes. The larger the distance between two different event nodes, the farther the distance is. Therefore, the relational feature distillation loss represents: Where D represents the mean square error of the distance measurement, and M represents the total number of events in the news text.

8. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1 is characterized by: In Step 3, the overall loss of the cross-language graph knowledge distillation model is expressed as: Among them, λ1 and λ2 are hyperparameters that control the loss weights of relational knowledge distillation and cross-view contrastive learning, and θ is all the parameters in the model that can participate in training, which is determined by λ 2Θ Regularization, using Adam as the optimizer during model training.

Images