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

Abstract

This invention relates to a cross-lingual news core event analysis method based on graph knowledge distillation, belonging to the field of natural language processing technology. Due to the scarcity of labeled data for news core event analysis in low-resource language scenarios, it is difficult to mine and model the global features and inter-event correlation features of news events in a text. To address this problem, this invention proposes a cross-lingual news core event analysis method based on graph knowledge distillation. This invention constructs an event hypergraph with event elements as vertices to represent global event features, and an event line graph with events as vertices to represent inter-event correlation features. Furthermore, it utilizes cross-lingual graph knowledge distillation to transfer event feature knowledge and relational knowledge from resource-rich languages ​​to low-resource languages, improving the effectiveness of news core event analysis in low-resource languages. Experimental results show that this method can effectively mine and model the global features and inter-event correlation features of news events in a text, and effectively alleviate the problem of scarce labeled data for news core event analysis in low-resource language scenarios.

Description

Technical Field

[0001] This invention relates to a cross-lingual news core event analysis method based on graph knowledge distillation, belonging to the field of natural language processing technology. Background Technology

[0002] News core analysis and detection technology quickly and accurately identifies key events and related information from news reports, providing users with an efficient way to obtain important information. Simultaneously, this technology greatly promotes in-depth analysis and data mining of news content, helping news organizations and researchers gain a deeper understanding of the background, impact, and relevance of events. Furthermore, it supports multilingual and cross-cultural information processing, providing a barrier-free information exchange platform for global users and enhancing user experience through personalized content recommendations. It also provides powerful auxiliary tools for news editors and reporters, improving work efficiency and reducing repetitive tasks. Ultimately, the application of news core analysis and detection technology not only improves the automation level of news processing but also drives innovation and development in the entire field of information dissemination.

[0003] Core news event analysis aims to analyze the importance of multiple related news events and identify those that best reflect the core theme of the news. Its core lies in effectively modeling the relationships between these events. However, due to the scarcity of labeled data for core news event analysis in low-resource language scenarios, mining and modeling the global features of news events and the relationships between them is challenging. To address this problem, a cross-lingual core news event analysis method based on graph knowledge distillation is proposed.

[0004] Effectively modeling the correlation features between events has been a hot research topic in the past. Previous researchers have used methods such as constructing core reference relationships, event frequency, and event location in news texts to alleviate these problems. However, these methods still have shortcomings in modeling the global semantic information of events. To overcome this limitation, Jindal et al. attempted to improve the performance of core event detection by strengthening event representation and encoding the potential interactions between events. Wang et al. used a question-answering model, integrating prior knowledge into the event extraction process, and identified core events by integrating similar events and scoring the extracted events. To model the correlations between events more deeply, Lu et al., based on the parts-of-speech and dependency relationships of words in the article, constructed complex structural relationships between words and sentences to detect core events.

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

[0006] (1) The global semantic information of an event cannot be effectively modeled by discrete event tuples and global event features alone;

[0007] (2) There are complex relationships between events in news texts, and attention mechanisms alone cannot effectively model the relationship 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] To address the aforementioned issues, this invention provides a cross-language news core event analysis method based on graph knowledge distillation. This invention effectively distills rich-resource knowledge to low-resource clients, improving the performance of low-resource language models.

[0010] This 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, improving the performance of low-resource language models. Specifically, addressing the aforementioned problems (1) and (2), a hypergraph structure is used to efficiently model the local and global semantic information of events. Specifically, firstly, an event hypergraph and an event line graph (ELG) are constructed based on the internal structure of events and the relationships between events. In the event hypergraph, a single event is regarded as a hyperedge, and its event elements are used as 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, hypergraph convolution and graph attention networks are used in the two channels of the event hypergraph and event line graph to obtain the event embeddings of each channel. Finally, cross-view self-supervised learning is introduced to learn the consistency of the event embeddings 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), a cross-lingual knowledge distillation framework is proposed based on hypergraph and line graph modeling. This framework utilizes prior knowledge learned from the resource-rich language (English) to compensate for the lack of labeled samples in the low-resource languages ​​(Chinese and Vietnamese). Specifically, the event feature knowledge obtained from the hypergraph channel and the relation feature knowledge obtained from the line graph channel are distilled into the corresponding hypergraph and line graph channels in the student model. Furthermore, cross-perspective self-supervised learning is introduced during the training of the student model to align the event embeddings of the hypergraph and line graph channels.

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

[0012] Step 1: Construct the event hypergraph and event line graph;

[0013] Step 2: Construct graph convolution, which 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 deep semantic information of events and structural information between events.

[0014] Step 3: Perform cross-language graph knowledge distillation using the 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 relation feature knowledge distillation module;

[0015] Cross-linguistic graph knowledge distillation is used to transfer knowledge from the resource-rich language teacher model to the low-resource language student model by utilizing 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 relational 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, thereby enriching the final event representation of low-resource language news text. Based on this representation, the coreness of events in news text can be further analyzed.

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

[0018] Step 1.1: Construct the event hypergraph:

[0019] Step 1.1.1: Represent the Event Hypergraph (EHG) as follows: ,in It is a node. It is a superedge, set the event as a superedge, and represent each superedge as... ,in Indicates the first The event mentioned in the event, Indicates the first The subject of the event, Indicates the first The object of the event, Indicates the first The time when the event occurred Indicates the first The location where each event occurred; each node is represented as... ,in For the first The first document in the document One event parameter;

[0020] Step 1.1.2, then use an association matrix. To represent the links between event nodes, where As an event element, Represents any hyperedge in the hypergraph;

[0021] Step 1.2: Construct an event line diagram:

[0022] Based on the event hypergraph, it is transformed into an event line graph, i.e., an ELG graph, to effectively model the interactions between event hyperedges and further model the deep semantic and structural information within and between events. The construction of the ELG graph is based on the following idea: hyperedges in the event hypergraph are treated as nodes in the line graph; two event nodes in the line graph are connected if and only if two corresponding event hyperedges share the same event feature. Given an event hypergraph... The line graph is represented as ,in Represents the vertices of a line graph. The hyperedges represent the hypergraphs of the line graph, which model the dependencies between events as a line graph of the event hypergraph;

[0023] Step 1.3: Based on the mBERT multilingual word embedding model, embedding of input text, EHG graph, and ELG graph nodes is achieved, mapping texts from different languages ​​to a unified semantic space; specifically:

[0024] For each token in the input text, a BERT embedding is obtained. Then, a bidirectional LSTM is used to encode the token of the event element for each event, generating a fixed-size event element representation. Finally, the element embeddings of all events are obtained. and event embedding It should be noted that during the training of the student model, mBERT is used to map texts from different languages ​​to the same semantic space.

[0025] Furthermore, Step 2 includes:

[0026] Step 2.1: Utilize hypergraph convolutional networks to capture high-order relationships between event elements, thereby obtaining event embeddings and modeling the complex relationships between events; define hypergraph convolution as:

[0027]

[0028] in, It is a non-linear activation function, including LeakyReLU or eLU. yes Hypergraph convolution of layers, It is the first For a specific node attribute in a layer, N1 represents the total number of event feature nodes in the event hypergraph, and M1 represents the total number of event hyperedges in the event hypergraph. It is the adjacency matrix of events; filters Used to extract features from nodes in the hypergraph, in the specific implementation, the weight of each hyperedge is set to 1, that is... After normalizing the above formula, it can be expressed as:

[0029]

[0030] in, Indicates the first The row normalization matrix of the layer, Represents the weight matrix. and These are diagonal matrices representing edge degree and vertex degree, respectively.

[0031] Hypergraph convolution efficiently captures high-order semantic relationships in an event hypergraph by performing a "node-hyperedge-node" feature transformation on the hypergraph; specifically, the multiplication operation... The information aggregation from vertex to hyperedge is defined, and then... Information aggregation from hyperedge to vertex was performed; initial event elements vertex features pass After hypergraph convolution, the node features obtained from each hypergraph convolution are averaged to obtain the final event element vertex embedding. The specific calculation formula is as follows:

[0032]

[0033] In the event hypergraph, event feature 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 feature embeddings, the hyperedge representation is obtained by mean pooling all these embeddings, thus obtaining the final event representation:

[0034]

[0035] in, Indicates the embedding of the event element hyperedge. It is a collection of event element nodes connected by hyperedges. This indicates the total number of event elements connected by the hyperedge. Is as the first Embedding of event element nodes in the input of the hypergraph convolutional layer.

[0036] Furthermore, Step 2 also includes:

[0037] Step 2.2: Based on hypergraph convolution, a line graph neural network is introduced to capture the interaction information between event hyperedges, further modeling the deep semantic and structural information of the event hypergraph; graph attention (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 event embedding; specifically, firstly, based on the weight matrix... , and Calculate the attention coefficient of the node:

[0038]

[0039] in, and Representing the nodes of the line graph and Embedded, Represents line graph nodes Line chart nodes The voting score, i.e., the line graph node Line chart nodes Attention coefficients are calculated by only computing nodes. of To introduce masked attention, where, Represents the nodes in the graph The neighborhood of the coefficients is used, and the softmax function is used to normalize the coefficients for easier computation.

[0040]

[0041] Finally, the weighted sum of the transition features with activation functions is calculated using the following formula:

[0042]

[0043] in, Represents the model's weight parameters. This 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 representation and generalization capabilities for event graphs. The specific calculation formula is shown below:

[0045]

[0046] in, This represents the total number of heads in a multi-head attention mechanism;

[0047] In each convolutional step, events aggregate information from their neighbors, embedding the initialized events into the input. In layered graph convolution, event-level information is captured; the final representation of the event embedding is calculated by averaging the embeddings obtained from each layer, and is expressed as:

[0048]

[0049]

[0050] in, In line graph convolution, the first... Layered event representation, This indicates the total number of convolutional layers in the line graph;

[0051] Step 2.3: The event hypergraph and event line graph are respectively processed... After convolution of the layer graphs, event embeddings from the perspective of the event hypergraph are obtained. Event embedding from the perspective of event line graph This leads to the final embedding of the event:

[0052]

[0053] in, It is a hyperparameter that balances the event representations of the two channels.

[0054] Furthermore, in Step 2, a cross-view self-supervised learning (CSSL) task is introduced to learn the consistency of event representations between the event hypergraph view and the event line graph view. In the specific implementation, the InfoNCE loss function is used to maximize the mutual information between these two views, thereby improving the model's performance in event representation learning.

[0055]

[0056] in, Measure the cosine similarity between two vectors. It is the total number of events in the document. 1 is the ordinal number of a specific event in the document. A temperature hyperparameter.

[0057] Furthermore, in Step 3, event feature knowledge distillation is first performed, specifically including:

[0058] Pre-trained cross-linguistic word embeddings are used to map low-resource language to the semantic space of rich-resource language. Specifically, this is first achieved by using temperature-sensitive embeddings... The Softmax function combines the event representations obtained from hypergraph modeling of the teacher model and the student model, respectively. Convert to soft tags This allows for the acquisition of more information within and between events, and also provides soft labels for the student model's event representations. ; For ease of writing, it is simplified to ;

[0059]

[0060] in, Soft labels representing event representations obtained from teacher-modeled event hypergraphs. and These represent event nodes in the event hypergraph channel. and event nodes Event representation, This represents the temperature coefficient hyperparameter, used to smooth the output of event features in the teacher model. Represents a node The set of adjacent nodes; the distillation loss of event feature knowledge distillation is expressed as:

[0061]

[0062] in, Represents the distance function. Indicates the total number of events in the news text;

[0063] Furthermore, to further improve the performance of the student model, real labels are used during model training; specifically, based on the final embeddings of the obtained events... Use a Softmax classification layer:

[0064]

[0065] in, and These are the learned parameters; the model training process involves using binary cross-entropy loss. :

[0066]

[0067] in, This represents the true label of the sample. It is a collection of real tags. This represents the predicted value of the sample. It is a set of predicted values; therefore, the event feature knowledge distillation loss can be written as a weighted average of the cross-entropy loss and the distillation loss:

[0068]

[0069] in, It is a hyperparameter that balances cross-entropy loss and distillation loss.

[0070] Furthermore, in Step 3, relational feature knowledge distillation is performed, specifically including:

[0071] This paper proposes an innovative knowledge distillation strategy, which transfers the relational knowledge learned in the teacher model's line graph channel to the corresponding line graph channel of the student model. The aim is 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.

[0072] Based on the line graph neural network, the first... Level Teacher Model Event Representation Event representation of student models A relational matrix is ​​designed to encode the knowledge of relationships between different event nodes, specifically represented as follows:

[0073]

[0074] in, Representing different event nodes, Used to measure the similarity between two event nodes. The larger the value, the greater the distance between two different event nodes. The same principle applies. Therefore, the relational feature distillation loss represents:

[0075]

[0076] in, This represents the mean square error of the distance measurement. This indicates the total number of events in the news text.

[0077] Furthermore, in Step 3, the overall loss of the cross-language graph knowledge distillation model is expressed as:

[0078]

[0079] in, and These are hyperparameters that control the weights of the loss for relational knowledge distillation and cross-perspective contrastive learning. For all parameters in the model that can participate in training, by Regularization, using Adam as the optimizer during model training.

[0080] The beneficial effects of this invention are:

[0081] 1. This invention provides a strategy for constructing an event hypergraph and an event line graph, and distills knowledge from a rich-resource model into a low-resource model through event feature knowledge distillation and relation feature knowledge distillation techniques, effectively alleviating the challenges when facing low-resource languages; this technology effectively improves the efficiency of core news event analysis;

[0082] 2. This invention was tested on three real-world datasets; compared with traditional baseline models, the experimental group showed significant improvements in evaluation metrics compared with all baseline models. Attached Figure Description

[0083] Figure 1 This is a schematic diagram illustrating the construction of the event hypergraph and event line graph of the present invention;

[0084] Figure 2 This is a diagram illustrating the cross-language graph knowledge distillation model architecture of the present invention.

[0085] Figure 3 This invention examines the impact of different model components on model performance on a Chinese dataset.

[0086] Figure 4 This invention examines the impact of different model components on model performance on a Vietnamese language dataset.

[0087] Figure 5 This invention examines the impact of the number of convolutional layers on model performance on a Chinese dataset.

[0088] Figure 6 This invention demonstrates the impact of the number of convolutional layers on model performance on a Vietnamese language dataset.

[0089] Figure 7 This invention examines the impact of hyperparameters on model performance on a Chinese dataset.

[0090] Figure 8 This invention examines the impact of hyperparameters on model performance on a Vietnamese language dataset. Detailed Implementation

[0091] Example 1: A cross-lingual news core event analysis method based on graph knowledge distillation, such as Figure 1 The diagram shown is a schematic representation of the overall event hypergraph and event line graph construction of the cross-language news core event analysis method based on graph knowledge distillation provided in this embodiment of the invention. Figure 2 This is a knowledge distillation model architecture diagram for a cross-language news core event analysis method based on graph knowledge distillation. The method includes the following steps:

[0092] Step 1: Construct an event hypergraph and event line graph. By analyzing the characteristics of events in news texts, it is found that events in news texts exhibit complex relationships such as dispersed event elements (event elements are distributed across different sentences or paragraphs) and overlapping event elements (multiple events share the same event element). Event hypergraphs and event line graphs can effectively model the complex interaction relationships within and between events, thereby capturing the local and global semantic information of events.

[0093] Furthermore, the specific steps of Step 1 include:

[0094] Step 1.1: Construct the event hypergraph:

[0095] Step 1.1.1: Represent the Event Hypergraph (EHG) as follows: ,in It is a node. It is a superedge, set the event as a superedge, and represent each superedge as... ,in Indicates the first The event mentioned in the event, Indicates the first The subject of the event, Indicates the first The object of the event, Indicates the first The time when the event occurred Indicates the first The location where each event occurred; each node is represented as... ,in For the first The first document in the document One event parameter;

[0096] Step 1.1.2, then use an association matrix. To represent the links between event nodes, where As an event element, Represents any hyperedge in the hypergraph;

[0097] like Figure 1 The event hypergraph described in the text, events The event elements included are ,event and events By sharing event elements This establishes a connection. It is important to further emphasize that the event hypergraph primarily captures high-order semantic relationships between event elements, thereby modeling the global semantic information of the event. To a certain extent, this also solves the problem of difficult semantic modeling of long texts in traditional text encoders.

[0098] Step 1.2: Construct an event line diagram:

[0099] Hypergraphs can model high-order semantic relationships and long-distance dependencies between event elements. However, events in news texts often exhibit complex relationships such as coreference, causality, temporality, and parent-child relationships. Modeling these relationship features is crucial for determining whether an event is a key event in the news text. Furthermore, research has demonstrated that line graphs can effectively model the implicit structural information between nodes in a graph. Therefore, this paper transforms the event hypergraph into an event line graph (ELG graph) to effectively model the interactions between event hyperedges and further model the deep semantic and structural information within and between events. The construction of the ELG graph is based on the following idea: hyperedges in the event hypergraph are treated as nodes in the line graph; two event nodes in the line graph are connected if and only if two corresponding event hyperedges share the same event element. Given an event hypergraph... The line graph is represented as ,in Represents the vertices of a line graph. The hyperedges represent the hypergraphs of the line graph, which model the dependencies between events as a line graph of the event hypergraph;

[0100] Step 1.3, Cross-linguistic Core Analysis and Detection: The first challenge to overcome is how to overcome the differences between different languages, eliminate the language gap, and achieve the representation and computation of multilingual news texts within a unified semantic space. Therefore, based on the mBERT multilingual word embedding model, the embedding of input text, EHG graph, and ELG graph nodes is implemented, mapping texts from different languages ​​to a unified semantic space; specifically:

[0101] For each token in the input text, a BERT embedding is obtained. Then, a bidirectional LSTM is used to encode the token of the event element for each event, generating a fixed-size event element representation. Finally, the element embeddings of all events are obtained. and event embedding It should be noted that during the training of the student model, mBERT is used to map texts from different languages ​​to the same semantic space.

[0102] Step 2: Construct graph convolution, which 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 deep semantic information of events and structural information between events.

[0103] Furthermore, Step 2 includes:

[0104] Step 2.1: Hypergraph convolutional networks are used 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:

[0105]

[0106] in, It is a non-linear activation function, including LeakyReLU or eLU. yes Hypergraph convolution of layers, It is the first For a specific node attribute in a layer, N1 represents the total number of event feature nodes in the event hypergraph, and M1 represents the total number of event hyperedges in the event hypergraph. It is the adjacency matrix of events; filters Used to extract features from nodes in the hypergraph, in the specific implementation, the weight of each hyperedge is set to 1, that is... After normalizing the above formula, it can be expressed as:

[0107]

[0108] in, Indicates the first The row normalization matrix of the layer, Represents the weight matrix. and These are diagonal matrices representing edge degree and vertex degree, respectively.

[0109] Hypergraph convolution efficiently captures high-order semantic relationships in an event hypergraph by performing a "node-hyperedge-node" feature transformation on the hypergraph; specifically, the multiplication operation... The information aggregation from vertex to hyperedge is defined, and then... Information aggregation from hyperedge to vertex was performed; initial event elements vertex features pass After hypergraph convolution, the node features obtained from each hypergraph convolution are averaged to obtain the final event element vertex embedding. The specific calculation formula is as follows:

[0110]

[0111] In the event hypergraph, event feature 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 feature embeddings, the hyperedge representation is obtained by mean pooling all these embeddings, thus obtaining the final event representation:

[0112]

[0113] in, Indicates the embedding of the event element hyperedge. It is a collection of event element nodes connected by hyperedges. This indicates the total number of event elements connected by the hyperedge. Is as the first Embedding of event element nodes in the input of the hypergraph convolutional layer.

[0114] Furthermore, Step 2 also includes:

[0115] Step 2.2: Event hypergraph modeling enables the model to effectively capture high-order semantic features between event elements. However, events described in news texts have strong contextual characteristics; each specific event establishes discourse relationships with other events in the news context. For example, Figure 1 In the same news text, the events described , and Events are linked by sharing the same event elements. Capturing the correlation features between events is crucial for modeling the core of news events. Therefore, a line graph neural network is introduced on top of hypergraph convolution to capture the interaction information between event hyperedges, further modeling the deep semantic and structural information of the event hypergraph. Another key to modeling the core of news events is capturing the interactive voting between events; events that receive higher votes from other events are assigned higher attention scores. To this end, a 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 event embedding; specifically, firstly, based on the weight matrix... , and Calculate the attention coefficient of the node:

[0116]

[0117] in, and Representing the nodes of the line graph and Embedded, Represents line graph nodes Line chart nodes The voting score, i.e., the line graph node Line chart nodes Attention coefficients are calculated by only computing nodes. of To introduce masked attention, where, Represents the nodes in the graph The neighborhood of the coefficients is used, and the softmax function is used to normalize the coefficients for easier computation.

[0118]

[0119] Finally, the weighted sum of the transition features with activation functions is calculated using the following formula:

[0120]

[0121] in, Represents the model's weight parameters. This represents the median weight matrix of the multi-head attention mechanism;

[0122] A multi-head attention mechanism is introduced to further enhance the model's representation and generalization capabilities for event graphs. The specific calculation formula is shown below:

[0123]

[0124] in, This represents the total number of heads in a multi-head attention mechanism;

[0125] In each convolutional step, events aggregate information from their neighbors, embedding the initialized events into the input. In layered graph convolution, event-level information is captured; the final representation of the event embedding is calculated by averaging the embeddings obtained from each layer, and is expressed as:

[0126]

[0127]

[0128] in, In line graph convolution, the first... Layered event representation, This indicates the total number of convolutional layers in the line graph;

[0129] Step 2.3: The event hypergraph and event line graph are respectively processed... After convolution of the layer graphs, event embeddings from the perspective of the event hypergraph are obtained. Event embedding from the perspective of event line graph This leads to the final embedding of the event:

[0130]

[0131] in, It is a hyperparameter that balances the event representations of the two channels.

[0132] Furthermore, in Step 2, although we represent the final event embedding as a synthesis of two views, the consistency between these two views is overlooked. Theoretically, to improve the model's generalization ability and robustness, we should guide the model to learn to ensure that the embedding representations of the same event are as close as possible in different hypergraph convolution channels, and vice versa. To address this limitation, a cross-view self-supervised learning (CSSL) task is introduced to learn the consistency of event representations between the event hypergraph view and the event line graph view. In the specific implementation, the InfoNCE loss function is used, aiming to maximize the mutual information between the two views, thereby improving the model's performance in event representation learning.

[0133]

[0134] in, Measure the cosine similarity between two vectors. It is the total number of events in the document. 1 is the ordinal number of a specific event in the document. A temperature hyperparameter.

[0135] Step 3: Perform cross-language graph knowledge distillation using the 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 relation feature knowledge distillation module;

[0136] Cross-linguistic graph knowledge distillation is used to transfer knowledge from the resource-rich language teacher model to the low-resource language student model by utilizing event feature knowledge distillation and relation feature knowledge distillation.

[0137] Specifically, the event feature knowledge extracted by the teacher model through the event hypergraph channel and the relational 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, thereby enriching the final event representation of low-resource language news text. Based on this representation, the coreness of events in news text can be further analyzed.

[0138] Furthermore, in Step 3, event feature knowledge distillation is first performed, specifically including:

[0139] Distilling event feature knowledge learned in rich-resource language hypergraph modeling into low-resource language can effectively improve the performance of news core analysis and detection models in low-resource language. Since the teacher and student models belong to different language modalities, the cross-linguistic problem needs to be addressed first. Therefore, pre-trained cross-linguistic word embeddings are used to map the low-resource language into the semantic space of the rich-resource language. Specifically, this is first achieved by using embeddings with temperature coefficients... The Softmax function combines the event representations obtained from hypergraph modeling of the teacher model and the student model, respectively. Convert to soft tags This allows for the acquisition of more information within and between events, and also provides soft labels for the student model's event representations. ; For ease of writing, it is simplified to ;

[0140]

[0141] in, Soft labels representing event representations obtained from teacher-modeled event hypergraphs. and These represent event nodes in the event hypergraph channel. and event nodes Event representation, This represents the temperature coefficient hyperparameter, used to smooth the output of event features in the teacher model. Represents a node The set of adjacent nodes; the distillation loss of event feature knowledge distillation is expressed as:

[0142]

[0143] in, Represents the distance function. Indicates the total number of events in the news text;

[0144] Furthermore, to further improve the performance of the student model, real labels are used during model training; specifically, based on the final embeddings of the obtained events... Use a Softmax classification layer:

[0145]

[0146] in, and These are the learned parameters; the model training process involves using binary cross-entropy loss. :

[0147]

[0148] in, This represents the true label of the sample. It is a collection of real tags. This represents the predicted value of the sample. It is a set of predicted values; therefore, the event feature knowledge distillation loss can be written as a weighted average of the cross-entropy loss and the distillation loss:

[0149]

[0150] in, It is a hyperparameter that balances cross-entropy loss and distillation loss.

[0151] Furthermore, in Step 3, relational feature knowledge distillation is performed, specifically including:

[0152] In the knowledge distillation process, focusing solely on transferring event-level knowledge to the student model is insufficient to effectively capture the complex interrelationships between events, which are crucial for accurately modeling event saliency. Through in-depth analysis, we observed that in the teacher model's line graph channel, the intermediate convolutional layers aggregate the relationships between embeddings of different event nodes. These relationships play a key role in modeling event saliency, directly revealing how the teacher model captures and models higher-order relationships between different nodes in the embedding feature space. Based on this, we propose an innovative knowledge distillation strategy: transferring the relational knowledge learned in the teacher model's line graph channel to the corresponding line graph channel of the student model. This aims to enhance the student model's ability to model global semantic information of events, thereby more accurately capturing and representing complex relationships between events. Specifically, an association matrix is ​​designed to encode the relationship knowledge between different event nodes from a pre-trained line graph neural network, and a type-related attention layer is introduced to automatically learn the importance of different event nodes.

[0153] Based on the line graph neural network, the first... Level Teacher Model Event Representation Event representation of student models A relational matrix is ​​designed to encode the knowledge of relationships between different event nodes, specifically represented as follows:

[0154]

[0155] in, Representing different event nodes, Used to measure the similarity between two event nodes. The larger the value, the greater the distance between two different event nodes. The same principle applies. Therefore, the relational feature distillation loss represents:

[0156]

[0157] in, This represents the mean square error of the distance measurement. This indicates the total number of events in the news text.

[0158] Furthermore, in Step 3, the overall loss of the cross-language graph knowledge distillation model is expressed as:

[0159]

[0160] in, and These are hyperparameters that control the weights of the loss for relational knowledge distillation and cross-perspective contrastive learning. For all parameters in the model that can participate in training, by Regularization, using Adam as the optimizer during model training.

[0161] To illustrate the effectiveness of this invention, it is compared with existing methods by testing on three datasets. The evaluation metric used is accuracy. Recall rate and standard recall rate ,in The calculation formula is as follows:

[0162]

[0163] Experiment 1: Comparative Experiment

[0164] Table 1. Experimental results of English, Chinese, and Vietnamese on the pre-trained teacher model.

[0165]

[0166] As shown in Table 1, the experimental results clearly demonstrate that the current state-of-the-art (SOTA) models KCE and CEE-IEA achieve significant performance on the English dataset in the news core analysis and detection task, with ACU values ​​as high as 69.76% and 73.49%, respectively. However, when these models are applied to the Chinese and Vietnamese datasets, although performance declines occur to varying degrees across all evaluation metrics, the baseline model CEE-IEA still achieves relatively respectable results in the AUC evaluation metric, at 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 with the corresponding experimental results in Tables 2 and 3, the results after fine-tuning the student model are better than those in Table 1. These experimental results strongly demonstrate the effectiveness of pre-trained teacher models in knowledge transfer; they can provide rich prior knowledge to the student model, thus providing possibilities and directions for improving the performance of news core analysis and detection models in low-resource language environments.

[0167] Compared to baseline models KCE and CEE-IEA, the model proposed in this invention achieves state-of-the-art results across all languages. Specifically, on the key evaluation metric AUC, this model achieves 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 of events and the correlation features between events.

[0168] Table 2. Performance comparison results of different baseline models on Chinese datasets

[0169]

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

[0171]

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

[0173] (1) In the evaluation of the baseline model (including Location, PageRank, and Frequency) which is entirely feature-based, we observed that Frequency and PageRank outperformed Location. This finding reveals the central role of event trigger words and event elements in the core news event, which are present throughout the entire report. 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, the neural network-based method significantly outperformed the feature-based method in overall performance. In particular, the KCE method demonstrated superior performance on all evaluation metrics, especially on P@1 and NR@1, where it achieved performance improvements of up to 10 percentage points, significantly surpassing benchmark models such as Location, PageRank, and Frequency. Furthermore, the experimental results also demonstrated the importance of modeling the correlation features between events 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, GAT-SED performed particularly well, demonstrating 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 was 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 syntax parsing function still needs improvement, thus leading to error cascading.

[0176] (4) In cross-linguistic scenarios, the GKDT proposed in this invention achieved the best experimental results among all baseline models. Specifically, in Chinese and Vietnamese environments, the AUC evaluation index was improved by 6.98% and 3.08%, respectively. Further comparison with the experimental results in Table 1 shows that the average performance improvement of the GKDT model in Chinese and Vietnamese is as high as 8.69% and 10.17%, respectively. These significant results verify the effectiveness of the event feature knowledge distillation and relation feature knowledge distillation strategies, which successfully transfer prior knowledge from the teacher model to the student model, thereby significantly improving the performance of low-resource language models. It is worth mentioning that GKDT performs better in Vietnamese than in Chinese scenarios, which fully demonstrates the strong competitiveness of the proposed method in low-resource language scenarios.

[0177] Experiment 2: Characteristic Ablation Experiment

[0178] To ensure the fairness and accuracy of the comparison, following the research paradigm of Jindal et al., we incorporated global event features into the model implementation. These global features cover 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 the aforementioned 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 on the evaluation metric AUC, and bolded numbers represent the best experimental results.

[0179] (1) w / o Parent Score (PS): Represents the number of child events that are not included in the model for a specific event.

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

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

[0182] (4) w / o Event Trigger Frequency (TF): indicates that events are not included in the model.

[0183] (5) w / o Argument Frequency (AF): indicates the frequency of occurrence of event elements not included in the model.

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

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

[0186]

[0187] The experimental results in Table 4 show that: (1) For both Chinese and Vietnamese, the performance of the model was significantly improved when global event features were incorporated into the model. Specifically, the Chinese and Vietnamese models achieved gains of 0.91% and 1.15% respectively on AUC, an important metric for evaluating model performance. This finding indicates that global features are crucial for improving the effectiveness of cross-language event analysis tasks. (2) In another evaluation metric, NR@k, as... As the values ​​increase, we find that the roles of TF (Frequency of Triggering Words) and AF (Frequency of Event Elements) features in the global event features also gradually strengthen. This observation further reveals the important value of TF and AF features in distinguishing salient events from other events, helping the model to more accurately identify and differentiate different events in complex event analysis tasks.

[0188] Experiment 3: Model Component Ablation Experiment

[0189] In the ablation experiment, the main focus was on analyzing the impact of different features and different module implementation structures on the performance of cross-language news core analysis and detection.

[0190] This invention primarily explores the impact of different component implementations in the GKDT model on the performance of cross-language news core analysis and detection. The proposed GKDT model is compared with the following variant models, specifically including:

[0191] (1) GKDT(-CL): indicates that the contrastive learning module is not used in the model to learn the consistency of event representations between the hypergraph channel and the line graph channel.

[0192] (2) GKDT(-ATT): indicates that GAT is not used in the line graph convolution in the model to model the correlation features between event nodes.

[0193] (3) GKDT(-EKD): This means that the event feature knowledge distillation component is removed during the model knowledge distillation stage, and only the cross-entropy loss in the student model is retained, that is... .

[0194] (4) GKDT(-RKD): indicates the removal of relation feature knowledge distillation components during the model knowledge distillation stage.

[0195] Depend on Figure 3 and Figure 4 From the experimental results, we can conclude that:

[0196] (1) In experiments with Chinese and Vietnamese corpora, the GKDT model proposed in this invention significantly outperformed all variant models in 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.

[0197] (2) During the training of the student model, we observed a decrease in model performance when contrastive learning and graph attention mechanisms were removed. This finding indicates that: firstly, contrastive learning plays a key role in capturing deep event semantic information and can effectively alleviate the problems caused by data imbalance and event hypergraph sparsity; secondly, graph attention mechanisms can effectively model the complex relationships between events and have an important impact on the task of analyzing core news events.

[0198] (3) During student model training, the removal of both event feature knowledge distillation (EKD) and relation feature knowledge distillation (RKD) significantly impacted model performance. This highlights the importance of simultaneously modeling EKD and RKD. By distilling the event feature knowledge and inter-event relation knowledge learned in the teacher model into the student model, the performance of the student model can be significantly improved.

[0199] (4) Comparing the two knowledge distillation methods (EKD and RKD), we found that EKD has a relatively limited effect on improving model performance, while RKD shows a more significant performance improvement, especially in the context of Vietnamese, a language with extremely low resource requirements. The above experimental results verify the transferability of inter-event correlation features and emphasize their important role in improving the performance of student models in low-resource language scenarios.

[0200] Experiment 4: Performance Analysis of the Number of Convolutional Layers

[0201] To delve deeper into the specific impact of different numbers of convolutional layers on model performance, this invention presents detailed experimental results of the GKDT model on Chinese and Vietnamese datasets, using different numbers of convolutional layers. Figure 5 and Figure 6As shown, on the Chinese dataset, when the GKDT model has 2 convolutional layers, it significantly outperforms other configurations across all evaluation metrics, demonstrating the best performance. Furthermore, 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 exhibits a drastically different trend. Specifically, when the number of convolutional layers is set to 3, the GKDT model demonstrates optimal performance, and the impact of different convolutional layer numbers on model performance is more significant. Further, as the number of convolutional layers increases, the performance of the GKDT model shows a significant downward trend across all evaluation metrics. A potential explanation for this phenomenon is that as the model depth increases, the GKDT model may be more susceptible to oversmoothing issues when aggregating high-order information from distant neighbors, leading to a gradual decline in its performance.

[0202] Experiment 5: Model Hyperparameter Analysis

[0203] Several hyperparameters were tuned on Chinese and Vietnamese datasets to validate the sensitivity of the GKDT model. More specifically, we used grid search to select the formula. Optimal hyperparameters The experimental results are as follows Figure 7 and Figure 8 As shown.

[0204] Depend on Figure 7 and Figure 8 From the experimental results, we can conclude that:

[0205] (1) Softmax temperature coefficient From the formula It can be seen that the temperature coefficient The higher the value, the smoother the probability distribution of event node categories. Figure 7 (a) and Figure 8 The experimental results in (a) show that, with the increase of the temperature coefficient... As the value gradually increases, the performance of the GKDT model corresponding to the method of this invention also shows a steady upward trend. In particular, when... When the temperature coefficient was set to 8, the model achieved optimal performance on multiple evaluation metrics, including NR@1, NR@5, NR@10, and AUC, demonstrating the positive impact of appropriately increasing the temperature coefficient on model performance. However, when the temperature coefficient... When the temperature coefficient exceeds 8, the model performance begins to show a slow downward trend. After analysis, we believe this may be due to excessive smoothing of the model caused by an excessively high temperature coefficient, which weakens its ability to distinguish different event node categories to some extent.

[0206] (2) Weight hyperparameters :formula The weight hyperparameters in the algorithm are used to balance the importance of cross-entropy loss and event feature knowledge distillation loss. Through... Figure 7 (b) and Figure 8 As can be observed from the experimental results in (b), we find that the GKDT model generally has better performance in terms of hyperparameters. The model is not particularly sensitive to changes. However, it is worth noting that on the Vietnamese dataset, the model... The sensitivity is enhanced compared to the Chinese dataset. Specifically, on the Chinese dataset, when... When set to 0.6, the model performance reaches its peak; however, on the Vietnamese dataset, The model performance reaches its optimal value of 0.8, and the performance fluctuates significantly as the value increases. This phenomenon not only verifies... This invention plays a crucial role in adjusting the model loss function and further confirms the effectiveness and applicability of the event feature knowledge distillation method proposed in low-resource language scenarios.

[0207] (3) Weight hyperparameters To further explore the specific impact of relational knowledge distillation on model performance, we systematically analyzed... Different weighting coefficients The performance of the model under the given values, such as Figure 7 (c) and Figure 8 As shown in (c). With As the value increases, the performance of the GKDT model shows an upward trend. Specifically, on the Chinese and Vietnamese datasets, when... When the values ​​were set to 100 and 110 respectively, the GKDT model achieved optimal performance. Further analysis revealed that on the Chinese dataset, when... When the values ​​are in the range [0.01, 1], the GKDT model exhibits relatively stable performance; however, on the Vietnamese dataset, the model performance decreases as the value falls within this range. The value increases and continues to strengthen. These experimental results fully demonstrate that the strategy of effectively distilling relational knowledge from the resource-rich language teacher model into the low-resource language student model significantly reduces the dependence of the news core event analysis model on labeled data.

[0208] (4) Weight hyperparameters In the formula The introduction of weight hyperparameters To control the magnitude of cross-view self-supervised learning, we designed a set of experiments to further explore the impact of cross-view self-supervised learning on model performance. The value range is set to [0.00, 0.01, 0.02, 0.03, 0.05], and in 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, when... When the value is set to 0.02, the GKDT model achieves optimal performance; however, on the Vietnamese dataset, The optimal value for 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 invention has significant advantages in low-resource language scenarios and can effectively reduce the model's dependence on a large amount of labeled data.

[0209] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A cross-lingual news core event analysis method based on graph knowledge distillation, characterized by: The specific steps of the method include: Step 1: Construct the event hypergraph and event line graph; Step 2: Construct graph convolution, which 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 deep semantic information of events and structural information between events. Step 3: Perform cross-language graph knowledge distillation using the 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 relation feature knowledge distillation module; Cross-linguistic graph knowledge distillation is used to transfer knowledge from the resource-rich language teacher model to the low-resource language student model by utilizing 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 relational 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, thereby enriching the final event representation of low-resource language news text. Based on this representation, the coreness of events in news text can be further analyzed. Step 2 includes: Step 2.1: Utilize hypergraph convolutional networks to capture high-order relationships between event elements, thereby obtaining event embeddings and modeling the complex relationships between events; define hypergraph convolution as: ； in, It is a non-linear activation function, including LeakyReLU or eLU. yes Hypergraph convolution of layers, It is the first For a specific node attribute in a layer, N1 represents the total number of event feature nodes in the event hypergraph, and M1 represents the total number of event hyperedges in the event hypergraph. It is the adjacency matrix of events; filters Used to extract features from nodes in the hypergraph, in the specific implementation, the weight of each hyperedge is set to 1, that is... After normalizing the above formula, it can be expressed as: ； in, Indicates the first The row normalization matrix of the layer, Represents the weight matrix. and These are 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 transformations on the hypergraph. Step 2.2: Based on hypergraph convolution, a line graph neural network is introduced to capture the interaction information between event hyperedges, further modeling the deep semantic and structural information of the event hypergraph; graph attention (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 event embedding; Step 2.3, the event hypergraph and event line graph are respectively processed... After convolution of the layer graphs, event embeddings from the perspective of the event hypergraph are obtained. Event embedding from the perspective of event line graph This leads to the final embedding of the event: ； in, It is a hyperparameter that balances the event representations of the two channels; Step 2 introduces a cross-view self-supervised learning (CSSL) task to learn the consistency of event representations between the event hypergraph view and the event line graph view. In its implementation, the InfoNCE loss function is used to maximize the mutual information between these two views, thereby improving the model's performance in event representation learning. ； in, Measure the cosine similarity between two vectors. It is the total number of events in the document. It is the ordinal number of a specific event in the document. A temperature hyperparameter.

2. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1, characterized in that: The specific steps of Step 1 include: Step 1.1: Construct the event hypergraph: Step 1.1.1: Represent the Event Hypergraph (EHG) as follows: ,in It is a node. It is a superedge, set the event as a superedge, and represent each superedge as... ,in Indicates the first The event mentioned in the event, Indicates the first The subject of the event, Indicates the first The object of an event. Indicates the first The time when the event occurred Indicates the first The location where each event occurred; each node is represented as... ,in For the first The first document in the document One event parameter; Step 1.1.2, then use an association matrix. To represent the links between event nodes, where As an event element, Represents any hyperedge in the hypergraph; Step 1.2: Construct an event line diagram: Based on the event hypergraph, it is transformed into an event line graph, i.e., an ELG graph, to effectively model the interactions between event hyperedges and further model the deep semantic and structural information within and between events. The construction of the ELG graph is based on the following idea: hyperedges in the event hypergraph are treated as nodes in the line graph; two event nodes in the line graph are connected if and only if two corresponding event hyperedges share the same event feature. Given an event hypergraph... The line graph is represented as ,in Represents the vertices of a line graph. The hyperedges represent the hypergraphs of the line graph, which model the dependencies between events as a line graph of the event hypergraph; Step 1.3: Based on the mBERT multilingual word embedding model, embedding of input text, EHG graph, and ELG graph nodes is achieved, mapping texts from different languages ​​to a unified semantic space; specifically: For each token in the input text, a BERT embedding is obtained. Then, a bidirectional LSTM is used to encode the token of the event element for each event, generating a fixed-size event element representation. Finally, the element embeddings of all events are obtained. and event embedding It should be noted that during the training of the student model, mBERT is used to map texts from different languages ​​to the same semantic space.

3. The cross-lingual news core event analysis method based on graph knowledge distillation according to claim 1, characterized in that: Step 2.1 includes: Specifically, multiplication operations The information aggregation from vertex to hyperedge is defined, and then... Information aggregation from hyperedge to vertex was performed; initial event elements vertex features pass After hypergraph convolution, the node features obtained from each hypergraph convolution are averaged to obtain the final event element vertex embedding. The specific calculation formula is as follows: ； In the event hypergraph, event feature 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 feature embeddings, the hyperedge representation is obtained by mean pooling all these embeddings, thus obtaining the final event representation: ； in, Indicates the embedding of the event element hyperedge. It is a collection of event element nodes connected by hyperedges. This indicates the total number of event elements connected by the hyperedge. Is as the first Embedding of event element nodes in the input of the hypergraph convolutional layer.

4. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1, characterized in that: Step 2.2 includes: Specifically, firstly, based on the weight matrix , and Calculate the attention coefficient of the node: ； in, and Representing the nodes of the line graph and Embedded, Represents line graph nodes Line chart nodes The voting score, i.e., the line graph node Line chart nodes Attention coefficients are calculated by only computing nodes. of To introduce masked attention, where, Represents the nodes in the graph The neighborhood of the coefficients is used, and the softmax function is used to normalize the coefficients for easier computation. ； Finally, the weighted sum of the transition features with activation functions is calculated using the following formula: ； in, Represents the model's weight parameters. This represents the median weight matrix of the multi-head attention mechanism; A multi-head attention mechanism is introduced to further enhance the model's representation and generalization capabilities for event graphs. The specific calculation formula is shown below: ； in, This represents the total number of heads in a multi-head attention mechanism; In each convolutional step, events aggregate information from their neighbors, embedding the initialized events into the input. In layered graph convolution, event-level information is captured; the final representation of the event embedding is calculated by averaging the embeddings obtained from each layer, and is expressed as: ； ； in, In line graph convolution, the first... Layered event representation, This indicates the total number of convolutional layers in the line graph.

5. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1, characterized in that: In Step 3, the event feature knowledge distillation is first performed, which specifically includes: Pre-trained cross-linguistic word embeddings are used to map low-resource language to the semantic space of rich-resource language. Specifically, this is first achieved by using temperature-sensitive embeddings... The Softmax function combines the event representations obtained from hypergraph modeling of the teacher model and the student model, respectively. Convert to soft tags This allows for the acquisition of more information within and between events, and also provides soft labels for the student model's event representations. ; For ease of writing, it is simplified to ; ； in, Soft labels representing event representations obtained from teacher-modeled event hypergraphs. and These represent event nodes in the event hypergraph channel. and event nodes Event representation, This represents the temperature coefficient hyperparameter, used to smooth the output of event features in the teacher model. Represents a node The set of adjacent nodes; the distillation loss of event feature knowledge distillation is expressed as: ； in, Represents the distance function. Indicates the total number of events in the news text; Furthermore, to further improve the performance of the student model, real labels are used during model training; specifically, based on the final embeddings of the obtained events... Use a Softmax classification layer: ； in, and These are the learned parameters; the model training process involves using binary cross-entropy loss. : ； in, This represents the true label of the sample. It is a collection of real tags. This represents the predicted value of the sample. It is a set of predicted values; therefore, the event feature knowledge distillation loss can be written as a weighted average of the cross-entropy loss and the distillation loss: ； in, It is a hyperparameter that balances cross-entropy loss and distillation loss.

6. The cross-language news core event analysis method based on graph knowledge distillation according to claim 1, characterized in that: In Step 3, relational feature knowledge distillation is performed, specifically including: This paper proposes an innovative knowledge distillation strategy, which transfers the relational knowledge learned in the teacher model's line graph channel to the corresponding line graph channel of the student model. The aim is 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. Based on the line graph neural network, the first... Level Teacher Model Event Representation Event representation of student models A relational matrix is ​​designed to encode the knowledge of relationships between different event nodes, specifically represented as follows: ； in, Representing different event nodes, Used to measure the similarity between two event nodes. The larger the value, the greater the distance between two different event nodes. The same principle applies. Therefore, the relational feature distillation loss represents: ； in, This represents the mean square error of the distance measurement. This indicates the total number of events in the news text.

7. The cross-lingual news core event analysis method based on graph knowledge distillation according to claim 1, characterized in that: In Step 3, the overall loss of the cross-language graph knowledge distillation model is expressed as: ； in, and These are hyperparameters that control the weights of the loss for relational knowledge distillation and cross-perspective contrastive learning. For all parameters in the model that can participate in training, by Regularization, using Adam as the optimizer during model training.

Images