A method for training a multi-document digest extraction model for public safety, a digest extraction method, and a device.

Abstract

This application provides a method, apparatus, and training method for summarizing multiple public safety documents. The method includes: generating a corresponding multi-document heterogeneous graph based on multiple original documents containing public safety content; training a local summarizing model using the multi-document heterogeneous graph, consisting of a graph convolutional network, a graph attention network, and a reinforcement learning model connected sequentially, so that the local summarizing model can output multiple target sentences corresponding to the input multi-document heterogeneous graph to form a summary data of the multi-document heterogeneous graph. This application can improve the efficiency and reliability of automatically acquiring cross-document information between multiple public safety documents, effectively improve the training efficiency and generalization of the public safety multi-document summarizing model, and thus effectively improve the comprehensiveness, accuracy, and efficiency of summarizing multiple public safety documents.

Description

Technical Field

[0001] This application relates to the field of document processing technology, and in particular to a method for training a summary extraction model for multiple documents in public security, a summary extraction method, and an apparatus. Background Technology

[0002] With the rapid growth and dissemination of data, information overload has become a common challenge. In the field of public safety, timely understanding and response to emergencies, early warning of potential risks, and analysis of abnormal activities all require the rapid and accurate summarization of large amounts of textual information. Document summarization provides an effective way to extract key information from documents and generate concise summaries that meet the diverse needs of users. For example, in the field of public safety, it can extract key information from a large number of news reports and social media messages, helping relevant personnel quickly understand the occurrence and development trends of events and take timely countermeasures. However, traditional methods, such as summarization based on word frequency and sentence similarity, are inefficient and lack accuracy when processing large-scale data.

[0003] Abstracting techniques can be broadly categorized into two types: extractive and generative. Extractive abstracting directly selects relevant sentences or phrases from the original document, then sorts and combines them to generate a summary. Therefore, extractive abstracting is more reliable grammatically and semantically, but may not be as fluent as generative abstracting. Generative abstracting, on the other hand, generates new sentences or paragraphs as a summary by semantically understanding and integrating the original document. Generative abstracting is more fluent, but may not be as reliable grammatically and semantically as extractive abstracting. Therefore, extractive abstracting is more effective in news reporting and scientific literature summarization, while generative abstracting is more effective in creative writing and emotional expression. Multi-document summarizing is an important branch of abstracting techniques, and it is more challenging and valuable than single-document summarizing. The task of multi-document summarizing is to create a concise summary of a collection of related documents. Compared to single-document summarizing, it needs to reduce redundancy while ensuring the integrity of the summary information, making it arguably more challenging. Extractive multi-document summarizing has significant application value in the field of public safety, as it can ensure the authenticity and readability of the summary.

[0004] However, due to the wide range of data sources for public safety documents, including public safety reports, monitoring logs, social media posts, and news articles, existing document summarization methods are not suitable for processing multi-source public safety documents. They also have difficulty automatically acquiring cross-document information and require significant time to correlate multiple public safety documents. Consequently, the accuracy and efficiency of summarization for multiple public safety documents cannot be guaranteed. Therefore, there is an urgent need to design an automatic multi-document summarization method that can ensure the comprehensiveness, accuracy, and efficiency of summarization for multiple public safety documents. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method for training a summary extraction model for multiple documents in public safety, a summary extraction method, and an apparatus to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of this application provides a method for training a summary extraction model for multiple documents in public safety, comprising:

[0007] A multi-document heterogeneous graph is generated based on multiple original documents containing public safety content. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association between the two nodes connected by the edge.

[0008] A local summary extraction model, consisting of a graph convolutional network, a graph attention network, and a reinforcement learning model connected sequentially, is trained using the multi-document heterogeneous graph. This local summary extraction model is used to output multiple target sentences corresponding to the input multi-document heterogeneous graph, thereby forming summary data of the multi-document heterogeneous graph.

[0009] In some embodiments of this application, the training of a local summarization extraction model consisting of a graph convolutional network, a graph attention network, and a reinforcement learning model using the multi-document heterogeneous graph, so that the local summarization extraction model can output multiple target sentences corresponding to the multi-document heterogeneous graph based on the input multi-document heterogeneous graph, thereby forming summary data of the multi-document heterogeneous graph, includes:

[0010] The multi-document heterogeneous graph is split into separate topological graphs, each centered on one of the sentence nodes.

[0011] A local summarization extraction model is trained based on each of the aforementioned topologies, consisting of sequentially connected graph convolutional networks, graph attention networks, and reinforcement learning models. The graph convolutional networks update the node representations of each input topology, outputting updated topologies for each topology. The graph attention networks then extract cross-document information from each sentence node in the updated topologies and update the sentence representations of each node accordingly, outputting the sentence representations corresponding to the multi-document heterogeneous graph. Finally, the reinforcement learning model, based on a pre-defined multi-step scenario Markov decision process strategy, extracts multiple target sentences corresponding to the multi-document heterogeneous graph through multiple time steps, forming the summary data for the multi-document heterogeneous graph.

[0012] In some embodiments of this application, after training the local summarization extraction model consisting of a graph convolutional network, a graph attention network, and a reinforcement learning model using the multi-document heterogeneous graph, the method further includes:

[0013] The trained local summarization extraction model is used as the local model parameters for the current iteration round. These local model parameters are then sent to a server belonging to the same federated learning system. The server receives these local model parameters and aggregates them with the local model parameters for the current iteration round sent by other client devices to obtain the global summarization extraction model parameters for the current iteration round. The server then distributes these global summarization extraction model parameters to each client device in the federated learning system as the local summarization extraction model to be trained by each client device in the next iteration round.

[0014] In some embodiments of this application, the non-sentence nodes include: document nodes, word nodes, and entity nodes; each document node is used to represent a different document; each sentence node is used to represent each sentence in the document in a one-to-one manner; each word node is used to represent each word in the sentence in a one-to-one manner; and each entity node is used to represent each public safety entity word in the document in a one-to-one manner.

[0015] Correspondingly, the association types of each node in the multi-document heterogeneous graph include: edges between different document nodes, edges between a document node and a sentence node, edges between a sentence node and a word node, edges between a sentence node and an entity node, edges between different sentence nodes, edges between different word nodes, and edges between different entity nodes.

[0016] In some embodiments of this application, generating a corresponding multi-document heterogeneous graph based on multiple original documents containing public safety content includes:

[0017] Each document containing public safety content is used as a different document node. Each sentence is extracted from each document and each sentence is used as a different sentence node. Each word is extracted from each sentence and each word is used as a different word node. Furthermore, each public safety entity term is extracted from each document and each public safety entity term is used as a different entity node.

[0018] Based on the document nodes, sentence nodes, word nodes, and entity nodes, as well as the relationships between different nodes, a multi-document heterogeneous graph corresponding to each document containing public safety content is constructed.

[0019] In some embodiments of this application, the reinforcement learning model is implemented based on an agent, which includes a global sentence encoder, a history extraction encoder, and an extractor connected in sequence.

[0020] The global sentence encoder is used to generate a global context information representation corresponding to each sentence representation for each sentence representation corresponding to the multi-document heterogeneous graph output by the graph attention network, and then send each sentence representation and the global context information representation corresponding to each sentence representation to the extraction history encoder and the extractor respectively.

[0021] The extraction history encoder is used to generate, in each time step, a historical information representation corresponding to each of the remaining sentences in the multi-document heterogeneous graph of the current time step (excluding the extracted target sentence), based on the sentence representation and the global context information representation corresponding to each of the remaining sentences, and send each of the historical information representations to the extractor.

[0022] The extractor is used to aggregate the sentence representation, global context information representation, and historical information representation corresponding to each of the remaining sentences in each time step to obtain the connection representation corresponding to each of the remaining sentences. Then, it calculates the score corresponding to each of the remaining sentences based on the connection representation corresponding to each of the remaining sentences, and obtains the summary extraction probability corresponding to the current time step based on a preset multi-head pooling layer. Then, it determines whether the summary extraction probability is greater than a preset probability threshold. If not, the remaining sentence with the highest score in the current time step is extracted and used as the extracted target sentence in the next time step. If so, the execution of the next time step is stopped, and each of the target sentences in the current time step is used as the summary data of the multi-document heterogeneous graph, and the summary data is output.

[0023] Another aspect of this application provides a method for summarizing multiple documents for public safety, including:

[0024] Receive a digest extraction request and retrieve multiple target documents containing public safety content specified in the digest extraction request;

[0025] Generate a multi-document heterogeneous graph corresponding to each of the target documents. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association relationship between the two nodes connected by the edge.

[0026] The multi-document heterogeneous graph is split into separate topological graphs, each centered on one of the sentence nodes.

[0027] Each of the aforementioned topological graphs is input into a local summary extraction model, so that the local summary extraction model outputs summary data of the multi-document heterogeneous graph corresponding to each of the target documents, wherein the local summary extraction model is pre-trained based on the aforementioned training method for summary extraction models of public safety multi-documents.

[0028] A third aspect of this application provides a training apparatus for a public safety multi-document summary extraction model, comprising:

[0029] The heterogeneous graph generation module is used to generate a corresponding multi-document heterogeneous graph based on multiple original documents containing public safety content. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association relationship between the two nodes connected by the edge.

[0030] The local model training module is used to train a local summary extraction model consisting of a graph convolutional network, a graph attention network, and a reinforcement learning model connected in sequence using the multi-document heterogeneous graph. This local summary extraction model is used to output multiple target sentences corresponding to the input multi-document heterogeneous graph to form summary data of the multi-document heterogeneous graph.

[0031] A fourth aspect of this application provides a summary extraction apparatus for multiple documents related to public safety, comprising:

[0032] The document receiving module is used to receive a digest extraction request and obtain multiple target documents containing public safety content specified in the digest extraction request.

[0033] A multi-document heterogeneous graph generation module is used to generate a multi-document heterogeneous graph corresponding to each of the target documents. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association relationship between the two nodes connected by the edge.

[0034] The topology graph partitioning module is used to split the multi-document heterogeneous graph into separate topology graphs, each centered on a sentence node.

[0035] The model extraction module is used to input each of the topological graphs into a local summary extraction model, so that the local summary extraction model outputs summary data of the multi-document heterogeneous graphs corresponding to each of the target documents. The local summary extraction model is pre-trained based on the aforementioned training method for summary extraction models of public safety multi-documents.

[0036] A fifth aspect of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the described method for training a summary extraction model for multiple public safety documents, and / or to implement the described method for extracting summaries for multiple public safety documents.

[0037] A sixth aspect of this application provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the described method for training a summary extraction model for multiple public security documents, and / or implements the described method for extracting summaries from multiple public security documents.

[0038] The seventh aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the described method for training a summary extraction model for multiple public security documents, and / or implements the described method for extracting summaries for multiple public security documents.

[0039] The eighth aspect of this application provides a federated learning system, comprising: a server, and various client devices respectively communicatively connected to the server;

[0040] The client device is used to execute the training method for the summary extraction model of multiple documents for public safety, and / or to execute the summary extraction method for multiple documents for public safety.

[0041] The client device is also used to send local model parameters in each iteration round to the server and receive global summary extraction model parameters sent by the server, so as to use the global summary extraction model parameters as the local summary extraction model to be trained in the next iteration round.

[0042] The server is configured to receive local model parameters sent by each of the client devices in each iteration round, aggregate the local model parameters of the current iteration round to obtain the global summary extraction model parameters of the current iteration round, and then send the global summary extraction model parameters to each of the client devices.

[0043] The method for training a public safety multi-document summary extraction model provided in this application generates a corresponding multi-document heterogeneous graph based on multiple original documents containing public safety content. The nodes in this multi-document heterogeneous graph belong to various semantic types, including sentence nodes and non-sentence nodes. Edges between nodes in the multi-document heterogeneous graph represent the association between the two nodes connected by the edge. A local summary extraction model, composed of a graph convolutional network, a graph attention network, and a reinforcement learning model connected sequentially, is trained using this multi-document heterogeneous graph. This local summary extraction model outputs multiple target sentences corresponding to the input multi-document heterogeneous graph to form summary data of the multi-document heterogeneous graph. This method improves the efficiency and reliability of automatically acquiring cross-document information between multiple public safety documents, effectively improves the training efficiency and generalization of the public safety multi-document summary extraction model, and thus effectively improves the comprehensiveness, accuracy, and efficiency of summarizing multiple public safety documents.

[0044] Additional advantages, objectives, and features of this application will be set forth in part in the description which follows, and will in part become apparent to those skilled in the art upon review of the following description, or may be learned by practice of the application. The objectives and other advantages of this application can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0045] Those skilled in the art will understand that the purposes and advantages that can be achieved with this application are not limited to those specifically described above, and that the above and other purposes that this application can achieve will be more clearly understood from the following detailed description. Attached Figure Description

[0046] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, do not constitute a limitation thereof. The components in the drawings are not drawn to scale but are merely for illustrating the principles of this application. For ease of illustration and description of certain parts of this application, corresponding portions in the drawings may be enlarged, i.e., may appear larger relative to other components in an exemplary device actually manufactured according to this application. In the drawings:

[0047] Figure 1 This is a schematic diagram of the first process of a method for training a summary extraction model for multiple documents for public safety, according to an embodiment of this application.

[0048] Figure 2 This is a schematic diagram of the second process of a method for training a summary extraction model for multiple documents for public safety, according to an embodiment of this application.

[0049] Figure 3 This is a schematic diagram illustrating documents, sentences, words, entities, and the relationships between them in one embodiment of this application.

[0050] Figure 4 This is a schematic diagram of the logical architecture of the training process of a summary extraction model for multiple documents for public safety in one embodiment of this application.

[0051] Figure 5 This is a flowchart illustrating a method for extracting summaries from multiple public security documents according to an embodiment of this application.

[0052] Figure 6 This is a schematic diagram of the structure of a training device for a public safety multi-document summary extraction model in one embodiment of this application.

[0053] Figure 7 This is a schematic diagram of a summary extraction device for multiple documents related to public safety according to an embodiment of this application. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and their descriptions are used to explain this application, but are not intended to limit it.

[0055] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the structures and / or processing steps closely related to the solution according to this application are shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0056] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0057] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0058] In the following description, embodiments of the present application will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0059] In one or more embodiments of this application, Graph Neural Networks (GNNs) are deep learning models used to process graph data. Graph data is a complex network structure composed of nodes and edges. The emergence of GNNs aims to effectively capture and utilize the structural information in graph data, thereby improving the learning of node or edge representations and task prediction. Reinforcement Learning (RL) is a machine learning paradigm whose goal is to enable an agent to learn through interaction with its environment, allowing the agent to take a series of actions in the environment to maximize the accumulated reward signal. Federated Learning is a distributed machine learning method that allows model training on devices or distributed nodes without sharing the original data. The goal of Federated Learning is to achieve global model training by updating model parameters locally and then aggregating the updated parameters on a central server, while protecting user privacy and data security. Multi-Document Summarization is a task in the field of natural language processing, whose goal is to extract key information from multiple documents and generate a concise and comprehensive summary. Unlike single-document summarization, multi-document summarization requires comprehensive consideration of the information relationships between multiple documents to ensure that the generated summary covers all important content.

[0060] First, to achieve summary extraction for multiple public safety documents (i.e., multiple public safety documents), it's possible to consider modeling the relationships between sentences based on graph structures. These methods can effectively capture overlapping information between documents when they all discuss the same topic. However, this approach fails if the content discussed between documents is complementary. Due to the differences between multiple public safety documents, it's necessary to design a graph structure that is applicable to multiple public safety documents and can effectively extract cross-document information.

[0061] Furthermore, Transformer-based pre-trained models (PLMs) have achieved significant results in various natural language processing (NLP) fields. Another popular approach is to design PLMs for text summarization. However, directly concatenating related documents into a sequence may fail to capture the relationships between documents. Moreover, input length limitations and computational complexity necessitate truncating documents during input, potentially resulting in the loss of crucial information. Therefore, a method is needed to effectively improve the generalization ability of summarization models for public safety multi-document processing.

[0062] Based on this, this application proposes a training method for a public safety multi-document summary extraction model that combines a multi-document heterogeneous graph and a local summary extraction model composed of sequentially connected graph convolutional networks, graph attention networks, and reinforcement learning models; a training device for the public safety multi-document summary extraction model that executes the training method; a public safety multi-document summary extraction method; a public safety multi-document summary extraction device for executing the method; a physical device; a computer-readable storage medium; and a computer program product. This method can improve the efficiency and reliability of automatically acquiring cross-document information between multiple public safety documents through the multi-document heterogeneous graph, and can improve the training effectiveness and generalization ability of the public safety multi-document summary extraction model through the local summary extraction model composed of sequentially connected graph convolutional networks, graph attention networks, and reinforcement learning models.

[0063] The following examples will provide a detailed description.

[0064] Based on this, embodiments of this application provide a method for training a summary extraction model for multiple public security documents, which can be implemented by a training device for a summary extraction model for multiple public security documents. See [link to relevant documentation]. Figure 1 The method for training a summary extraction model for multiple documents in public safety specifically includes the following:

[0065] Step 100: Generate a corresponding multi-document heterogeneous graph based on multiple original documents containing public safety content. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association relationship between the two nodes connected by the edge.

[0066] Specifically, step 100 models the document as a heterogeneous graph, and by introducing multiple semantic nodes to enrich the relationships between sentences, it can improve the efficiency and reliability of automatically acquiring cross-document information between multiple public security documents.

[0067] In one or more embodiments of this application, the multiple original documents containing public safety content may include public safety reports, monitoring logs, social media posts, and news articles, etc.

[0068] Step 200: Train a local summary extraction model using the multi-document heterogeneous graph, which consists of a graph convolutional network, a graph attention network, and a reinforcement learning model connected in sequence, so that the local summary extraction model can output multiple target sentences corresponding to the multi-document heterogeneous graph based on the input multi-document heterogeneous graph, so as to form the summary data of the multi-document heterogeneous graph.

[0069] Specifically, Graph Attention Networks (GAT) and Graph Convolutional Networks (GCN) are used to distinguish the relationships between different types of semantic nodes, which helps extract semantic representations across documents. However, extracting summaries solely based on node probabilities may lead to repetitive summary information because it ignores the attention paid to previously extracted sentences. Furthermore, node probability-based methods require different probability thresholds for different datasets, which limits the model's generalization ability. Therefore, to address this issue, step 200 of this embodiment also employs a reinforcement learning model. This model considers the historical information of summary extraction at each time step (also referred to as a time step) and the overall information of the document to determine whether the network output should continue to use sentences as summaries. This effectively reduces the redundancy of summary information and significantly improves the training effectiveness and generalization ability of the public safety multi-document summary extraction model.

[0070] As can be seen from the above description, the method for training a summary extraction model for multiple public safety documents provided in this application can improve the efficiency and reliability of automatically acquiring cross-document information between multiple public safety documents, effectively improve the training efficiency and generalization of the summary extraction model for multiple public safety documents, and thus effectively improve the comprehensiveness, accuracy and efficiency of the summary extraction of multiple public safety documents.

[0071] To further improve the effectiveness and reliability of training a local summarization extraction model composed of sequentially connected graph convolutional networks, graph attention networks, and reinforcement learning models using the aforementioned multi-document heterogeneous graph, a method for training a summarization extraction model for public security multi-document applications is provided in this application embodiment. (See also...) Figure 2 Step 200 in the method for training a summary extraction model for multiple documents in public safety specifically includes the following:

[0072] Step 210: Split the multi-document heterogeneous graph into individual topological graphs, each centered on a sentence node.

[0073] Step 220: Train a local summarization extraction model based on each of the topological graphs, consisting of a graph convolutional network, a graph attention network, and a reinforcement learning model connected sequentially. The graph convolutional network updates the node representations of each input topological graph to output the updated topological graphs. Then, the graph attention network extracts cross-document information from each sentence node in the updated topological graphs and updates the sentence representations of each sentence node accordingly, outputting the sentence representations corresponding to the multi-document heterogeneous graph. Finally, the reinforcement learning model, based on a preset multi-step scenario Markov decision process strategy, extracts multiple target sentences corresponding to the multi-document heterogeneous graph through multiple time steps according to the input sentence representations to form the summary data of the multi-document heterogeneous graph.

[0074] Specifically, the multi-step scenario Markov decision process strategy refers to the multi-step scenario Markov decision process (MDP), which can be implemented using MemSum (a multi-step scenario Markov decision process extractor and summarizer). By considering the historical information extracted from the summary at each time step and the overall information of the document, the decision network outputs whether to continue using the sentence as a summary, which can effectively reduce the redundancy of summary information.

[0075] Understandably, this application combines the advantages of graph neural networks, reinforcement learning, and federated learning to address the unique challenges of public safety data. The public safety field typically involves massive amounts of text data from diverse sources, including public safety reports, monitoring logs, social media posts, and news articles. This data is often characterized by high timeliness, high information sensitivity, large volume, and strong heterogeneity. Therefore, using traditional recurrent neural networks for summarization can lead to inaccurate and inefficient information processing. When processing monitoring logs or social media posts, due to the real-time and high-frequency update characteristics of the data, directly using large models for summarization is not only computationally complex, but the limitations of the large models themselves may also lead to errors in the summarization results, misleading users. Therefore, to improve the computational efficiency and comprehensiveness of summaries for public safety data, and to improve the accuracy and efficiency of processing public safety data, this application uses graph neural networks to model the data, enabling the capture of complex relationships and structural features between massive amounts of data. Documents are modeled as heterogeneous graphs, and various semantic nodes are introduced to enrich the relationships between sentences.

[0076] In a specific example, graph convolutional networks and graph attention networks are used to update node representations in a multi-document heterogeneous graph (heterogeneous graph for short). For the graph convolutional network, given a heterogeneous graph G and node features at layer l... Indicates a dimension of d h The real vector space. A graph convolutional network can be represented as:

[0077]

[0078] in This represents the feature representation of node v at layer (l+1). At each layer, the state of a node is calculated by summarizing information from its neighbors. AGG r∈R Let R represent an aggregation function that summarizes information from neighboring nodes onto node v. Here, R represents the set of neighboring relations. It is an update function used to calculate information about the relationship r between node u and node v. It is the feature representation of node u in layer l.

[0079] The update design for a node with relation r is as follows:

[0080]

[0081] Where R is the set of all relation types in the heterogeneous graph. It is the feature representation of node v in the relational subgraph r at layer l+1. This is the feature representation of node u in layer l within the relational subgraph r. σ is the ReLU activation function. ij It is a constant that normalizes the contributions of neighboring nodes. f is the set of neighboring nodes of node v under relation r. r It is a convolution function for relation type r. r It is a subgraph with relation r. It is the weight matrix of the l-th convolutional layer with relation r. Let be the bias vector associated with relation type r in layer l. During node update, different relation representations of the same node are concatenated to form the node representation for the next layer. That is:

[0082]

[0083] in Let v be the feature representation of node v in the subgraph with relation r at layer l. || denotes the join vector operation. Taking sentence node updates as an example, by utilizing graph convolutional networks, the structural features of heterogeneous graphs are captured, thereby obtaining the semantic node representations of the entire document cluster. Simultaneously, through join operations, feature information of all relation types is preserved to enhance the model's expressive power.

[0084] For graph attention networks, since summary extraction is performed on a sentence-by-sentence basis, the focus is on using graph attention networks to update sentence nodes. Considering the relationship between sentence nodes and other semantic nodes, the heterogeneous graph attention network is designed as follows:

[0085]

[0086]

[0087]

[0088] Among them W r,T W l These are the trainable parameters of a graph attention network. This represents the attention score between node s and its neighbor node o in the l-th layer of the relation subgraph r. Let represent the attention scores of node s and its k-type neighbor nodes in the l-th layer of the relation subgraph r. LeakyReLU is represented as max(αx,x) where α is a negative slope less than 1. Let r represent the set of neighbors of node s with relation r. It is the attention coefficient between sentence nodes and other semantic nodes (including sentence nodes). This represents the feature representation of a sentence node at layer l under the relation subgraph r. α i,j It's the attention weight. W lThese are the learnable weights of the l-th layer. N represents the feature representation of nodes other than sentence nodes at layer l under the relation subgraph r. i It is the set of neighboring nodes of a sentence node. The multi-head attention mechanism is described as follows: The attention coefficients and weight matrix are expanded into multiple heads, each calculated independently, and the results are concatenated.

[0089]

[0090] in W is the semantic coefficient between the sentence node and other semantic nodes (including the sentence node) at the h head. l,h These are the trainable parameters of the h-th head in the l-th layer. Let O be the feature representation of node o in the l-th layer.

[0091] The final sentence representation is obtained by concatenating the sentence representations updated based on all relations:

[0092]

[0093] in It is the feature representation of the sentence node in the (l+1)th layer.

[0094] To address the issue of cross-platform data incompatibility, further ensure the independence and security of various data sources, effectively prevent data leakage and unauthorized access, and maintain data privacy and integrity, this application provides a method for training a public security multi-document summary extraction model, see [link to relevant documentation]. Figure 2 The method for training a public safety multi-document summary extraction model, after step 200, further includes the following:

[0095] Step 300: The trained local summarization extraction model is used as the local model parameter for the current iteration round, and the local model parameter is sent to a server belonging to the same federated learning system. The server receives the local model parameter and aggregates it with the local model parameters for the current iteration round sent by other client devices to obtain the global summarization extraction model parameter for the current iteration round. The server then distributes the global summarization extraction model parameter to each client device in the federated learning system as the local summarization extraction model to be trained by each client device in the next iteration round.

[0096] Based on this, the embodiments of this application can not only extract summaries quickly and efficiently to effectively reduce the redundancy of summary information, but also solve the problem of cross-platform data incompatibility by combining federated learning.

[0097] In other words, to address the issue of cross-platform data incompatibility, this application's embodiments incorporate federated learning. By processing and training data locally on various devices or nodes, it ensures that data integration does not compromise data privacy or integrity. Specifically, federated learning allows data processing and model training to be completed at the data source end, transmitting only encrypted model parameters without transmitting the original data. This not only guarantees the independence and security of each data source but also effectively prevents data leakage and unauthorized access, maintaining data privacy and integrity.

[0098] Entities contain rich semantic information, which can be seen as markers of important information, crucial for summarizing and reducing redundancy. Existing graph structures primarily focus on relationships between sentences and words. Therefore, to further capture cross-document information between sentences, different attention weights are assigned based on the relationships between nodes to capture important information. In the public security multi-document summarization extraction model training method provided in this application embodiment, the non-sentence nodes specifically include the following:

[0099] Document nodes, word nodes, and entity nodes; each document node represents a different document; each sentence node represents a sentence in a document in a one-to-one manner; each word node represents a word in a sentence in a one-to-one manner; and each entity node represents a public safety entity in the document in a one-to-one manner.

[0100] Correspondingly, the association types of each node in the multi-document heterogeneous graph include: edges between different document nodes, edges between a document node and a sentence node, edges between a sentence node and a word node, edges between a sentence node and an entity node, edges between different sentence nodes, edges between different word nodes, and edges between different entity nodes.

[0101] Specifically, multi-sample heterogeneous graphs contain four types of nodes V s V w V e V d and 7 kinds of edges e dd ,e ds ,e sw ,e se ,e ss ,e ww ,e ee Among them, V s V w V eV d These represent sentence nodes, word nodes, entity nodes, and document nodes, respectively. dd ,e ds ,e sw ,e se ,e ss ,e ww ,e ee The edges are represented sequentially as follows: edges between different document nodes, edges between a document node and a sentence node, edges between a sentence node and a word node, edges between a sentence node and an entity node, edges between different sentence nodes, edges between different word nodes, and edges between different entity nodes.

[0102] In one or more embodiments of this application, the entity refers to words such as names of people, organizations, locations, times, and products in the document.

[0103] In other words, this application uses heterogeneous graphs to model public safety document data. The heterogeneous graph consists of four types of semantic nodes: documents, sentences, words, and entities. Then, in step 200, high-level hidden features are generated for each semantic node of the heterogeneous graph using GCN, and the semantic relationships between these nodes are encoded into the hidden features of the nodes. Cross-document information between sentences is captured using GAT, and different attention weights are assigned according to the relationships between nodes, thereby capturing important information.

[0104] Based on this, in the summary extraction model training method for multiple documents in public security provided in the embodiments of this application, see... Figure 2 Step 100 of the method for training a summary extraction model for multiple documents in public safety specifically includes the following:

[0105] Step 110: Each document containing public safety content is treated as a different document node; each sentence is extracted from each document and treated as a different sentence node; each word is extracted from each sentence and treated as a different word node; and each public safety entity term is extracted from each document and treated as a different entity node.

[0106] See Figure 3 and Figure 4A document contains sentences, and sentences contain words and entities. Here, w represents a word node; e represents an entity node; S represents a sentence node; and d represents a document node. d1, d2, and d3 represent different document nodes; S1 to S3 represent different sentence nodes; and e1 to e5 represent different entity nodes.

[0107] Step 120: Based on the document nodes, sentence nodes, word nodes, and entity nodes, as well as the relationships between different nodes, construct a multi-document heterogeneous graph corresponding to each document containing public safety content.

[0108] In a specific example, to effectively model documents and capture the relationships between semantic units at different levels, this application constructs a multi-document heterogeneous graph G = {V, E} to represent document clusters. Where V = V d ∪V s ∪V w ∪V e It is a set of nodes, E = e dd ∪e ds ∪e sw ∪e se ∪e ss ∪e ww ∪e ee It is an edge set. First, each document is represented as a document node V in the heterogeneous graph. d Then, generate a sentence node V for each sentence in the document. s And generate a word node V for each unique word. w In addition, to enrich the semantic nodes, this application also adds entity node V. e This hierarchical structure helps capture contextual information at different granularities within a document.

[0109] Next, seven types of edges are introduced. These are the edges between documents. dd The edge between document and sentence ds The border between sentences and words sw The edge e between sentences and entities se The border between sentences ss The edges between words ww The edge between entities ee Besides e ss All edges are undirected, in order to obtain the contextual semantics of the sentence. ssThe sentences in the document are sequentially connected. This forms a connected graph where, after multiple layers of message passing, each semantic node receives messages from its neighbors and updates its representation. A pre-trained model is used to obtain word embeddings, entity embeddings, and sentence embeddings (document embeddings use the average of the sentence embeddings) to initialize the nodes of the heterogeneous graph.

[0110] To leverage reinforcement learning, this application considers extracting historical information and overall document information during iterative selection of summary sentences, reducing redundancy in summaries and enhancing the model's generalization ability across different datasets. In a training method for a public safety multi-document summary extraction model provided in this embodiment, see [link to relevant documentation]. Figure 4 The original documents consist of public safety-related text data. During federated learning, the aggregation function uses the FedAvg method: i is aggregated. Where w t Let w' be the global parameters of the model in round t. n is the number of clients participating in training. t-1,i Here are the model parameters for client i in round t-1. The reinforcement learning model in the training method for summarizing multiple documents for public safety is implemented based on an agent, which includes a global sentence encoder GlobCE, an extraction history encoder EHE, and an extractor connected sequentially.

[0111] The global sentence encoder is used to process each sentence representation g corresponding to the multi-document heterogeneous graph output by the graph attention network. si Each of the above sentences generates its corresponding global context information representation C. si Then, the representations of each sentence and the corresponding global context information of each sentence representation are sent to the extraction history encoder and the extractor, respectively.

[0112] The extraction history encoder is used, at each time step, to extract the remaining sentences (excluding the extracted target sentence) from the multi-document heterogeneous graph in the current time step, based on the sentence representation gsi and the global context information representation C corresponding to each remaining sentence. si Generate historical information representation h for each remaining sentence in the current time step. si (By combining the extracted sentences and the unextracted sentences, we obtain historical information representations), and then assign each of the historical information representations h... si Send to the extractor;

[0113] The extractor is used to extract the sentence representation g corresponding to each of the remaining sentences at each time step. si The global context information represents C. si The historical information representation hsi is aggregated to obtain the connection representation corresponding to each of the remaining sentences. Then, the score corresponding to each of the remaining sentences is calculated based on the connection representation corresponding to each of the remaining sentences. (That is: the final score of each remaining sentence after passing through the extractor, the one with the highest score is selected as the summary content), and the summary extraction probability p corresponding to the current time step is obtained based on the preset multi-head pooling layer MHP. stop (That is, the probability of whether to continue extracting the summary), and then determine whether the probability of extracting the summary is greater than the preset probability threshold p. threshold If not, the remaining sentence with the highest score in the current time step is extracted to obtain a sample sentence and used as the extracted target sentence for the next time step; if yes, the next time step is stopped, and each target sentence in the current time step is used as the summary data of the multi-document heterogeneous graph, and the summary data is output, where the Compute Reward refers to the arithmetic mean of the ROUGE-1, ROUGE-2 and ROUGE-L scores.

[0114] Specifically, the policy gradient method is one of the commonly used algorithms in reinforcement learning. It continuously adjusts policy parameters to make the agent more likely to choose actions that yield high rewards. In the summary extraction process, this application employs the policy gradient method to find the optimal policy and maximize summary quality. During iterative summary extraction, the reward function represents the overall quality of the summary, typically calculated using metrics such as ROUGE. The reward is calculated as the arithmetic mean of the ROUGE-1, ROUGE-2, and ROUGE-L scores.

[0115] reward=(ROUGE-1+ROUGE-2+ROUGE-L) / 3 (9)

[0116] The policy gradient is calculated as the expected value of the total reward multiplied by the logarithmic gradient of the probability of the action chosen according to the policy.

[0117]

[0118]

[0119] Where J(θ) represents the expected reward function under policy parameter θ, and π(At|St,θ) represents the given state S t Strategy πθ Select action A t The probability of.

[0120] Based on a multi-step contextual MDP strategy, the abstract extraction consists of multiple time steps. At each time step t, the action is to either stop or select a sentence for the candidate abstract. The strategy is described as follows:

[0121] π(At|St,θt)=p(stop|St,θt)p(at|stop,St,θt) (12)

[0122]

[0123] Where π(At|St,θt) represents the policy function in state S t Next. Perform a certain action A. t The probability of stopping in state S. Θt is the policy network parameter. p(stop|St,θt) represents the probability of stopping in state S. t And the probability of deciding whether to stop, given the parameter Θt. p(at|stop,St,θt) represents the action a to be chosen after deciding to stop. t The probability of I. t This represents the set of remaining sentence indices at time step t. j It is the score of the remaining sentences. To perform action a t The score of the selected sentence.

[0124] At time step t, the AI ​​decides whether to stop abstract extraction based on the strategy.

[0125] The Global Context Encoder (GCE) consists of an N g It consists of layers of bi-LSTM, which takes L local sentence embeddings as input and generates a global contextual information representation for each sentence. Should Encode global contextual information, including information such as the sentence's position in the document and information about adjacent sentences.

[0126] The Extraction History Encoder (EHE) encodes the extraction history information and generates an extraction history embedding for each remaining sentence. The Extraction History Embedding (EHE) consists of a stack of n identical layers. Within each layer, there are two multi-head attention sublayers. One sublayer performs multi-head self-attention (MHA) between local embeddings, ensuring that each remaining sentence captures the context provided by the other remaining sentences. The other attention sublayer performs multi-head attention on the embeddings of the extracted sentences, ensuring that each remaining sentence pays attention to all extracted sentences. The outputs of the two attention sublayers, one for each remaining sentence, capture the contextual information of both the extracted and remaining sentences. The final output of the EHE constitutes the extraction history embeddings, one for each remaining sentence.

[0127] The extractor calculates a score for each remaining sentence and outputs an extraction stop signal. As input to the extractor, this application generates an aggregation layer for each remaining sentence by concatenating local sentence embeddings, global context embeddings, and extraction history embeddings. To generate sentence scores, the concatenated embeddings of the remaining sentences are passed to a fully connected layer via ReLU activation, and then projected as a scalar through the fully connected layer and a sigmoid function, where the sigmoid function is denoted as... The same fully connected layer is applied to all remaining sentences. The extractor can learn to stop extracting based on the state of the remaining sentences. Therefore, MHP is applied to the last hidden vector of all remaining sentences to output a single vector. This vector is then passed to a linear layer with a sigmoid function to produce the summary extraction probability p. stop If p stop If the probability exceeds this threshold, stop extracting the summary sentence; if it is less than or equal to this threshold, continue extracting the summary sentence.

[0128] The input to the global sentence encoder is... That is, the sentence representation obtained after processing by the graph encoder. The output is... This represents the global context information of the sentence. It is understood that in one or more embodiments of this application, the representation refers to a vector; for example, the sentence representation is a sentence vector.

[0129] In summary, this application provides a method for training a summary extraction model for multiple documents in public safety. It generates high-level hidden features for each semantic node in a heterogeneous graph using a graph convolutional network and encodes the semantic relationships between these nodes into these hidden features. A graph attention network captures cross-document information between sentences, assigning different attention weights based on the relationships between nodes to capture important information. Furthermore, by designing a reinforcement learning module that considers extracting historical information and overall document information during iterative selection of summary sentences, redundancy in the summary is significantly reduced, and the model's generalization ability across different datasets is enhanced. Simultaneously, federated learning addresses the issue of data incompatibility between different platforms.

[0130] Based on the summary extraction model training method for multiple public security documents provided in the foregoing embodiments, this application also provides an embodiment of a summary extraction method for multiple public security documents. This summary extraction method for multiple public security documents can be implemented by a summary extraction device for multiple public security documents. See [link to relevant documentation]. Figure 5 The method for extracting summaries from multiple public safety documents specifically includes the following:

[0131] Step 400: Receive a digest extraction request and obtain multiple target documents containing public safety content specified in the digest extraction request.

[0132] Step 500: Generate a multi-document heterogeneous graph corresponding to each of the target documents. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association relationship between the two nodes connected by the edge.

[0133] Step 600: Split the multi-document heterogeneous graph into separate topological graphs, each centered on a sentence node.

[0134] Step 700: Input each of the topology graphs into the local summary extraction model so that the local summary extraction model outputs summary data of the multi-document heterogeneous graphs corresponding to each of the target documents, wherein the local summary extraction model is pre-trained based on the aforementioned training method for the summary extraction model of public safety multi-documents.

[0135] As can be seen from the above description, the summary extraction method for multiple public safety documents provided in this application can improve the efficiency and reliability of automatically obtaining cross-document information between multiple public safety documents, thereby effectively improving the comprehensiveness, accuracy and efficiency of summary extraction from multiple public safety documents.

[0136] From a software perspective, this application also provides a training apparatus for a public security multi-document digest extraction model, comprising all or part of the aforementioned training method for training public security multi-document digest extraction models. See [link to relevant documentation]. Figure 6 The training device for the public safety multi-document summary extraction model specifically includes the following components:

[0137] The heterogeneous graph generation module 10 is used to generate a corresponding multi-document heterogeneous graph based on multiple original documents containing public safety content. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association relationship between the two nodes connected by the edge.

[0138] The local model training module 20 is used to train a local summary extraction model consisting of a graph convolutional network, a graph attention network, and a reinforcement learning model connected in sequence using the multi-document heterogeneous graph, so that the local summary extraction model can output multiple target sentences corresponding to the multi-document heterogeneous graph based on the input multi-document heterogeneous graph, so as to form the summary data of the multi-document heterogeneous graph.

[0139] The embodiments of the public security multi-document summary extraction model training device provided in this application can be used to execute the processing flow of the public security multi-document summary extraction model training method embodiments described above. Its functions will not be repeated here, but can be referred to the detailed description of the public security multi-document summary extraction model training method embodiments described above.

[0140] The portion of the training device for the public security multi-document summarization model can be performed on the client device. The specific implementation can be chosen based on the processing power of the client device and limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are performed on the client device, the client device may further include a processor for specific processing of the public security multi-document summarization model training.

[0141] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0142] The server and the client device can communicate using any suitable network protocol, including those not yet developed as of the date of this application. Such network protocols may include, for example, TCP / IP, UDP / IP, HTTP, HTTPS, etc. Furthermore, such network protocols may also include RPC (Remote Procedure Call Protocol) and REST (Representational State Transfer Protocol) protocols used on top of the aforementioned protocols.

[0143] As can be seen from the above description, the training device for the summary extraction model of multiple public safety documents provided in this application can improve the efficiency and reliability of automatically acquiring cross-document information between multiple public safety documents, effectively improve the training efficiency and generalization of the summary extraction model of multiple public safety documents, and thus effectively improve the comprehensiveness, accuracy and efficiency of the summary extraction of multiple public safety documents.

[0144] To address the issue of cross-platform data incompatibility, further ensure the independence and security of various data sources, effectively prevent data leakage and unauthorized access, and maintain data privacy and integrity, this application provides a training device for a public security multi-document summary extraction model, see [link to relevant documentation]. Figure 6 The training device for the public safety multi-document summary extraction model also specifically includes the following:

[0145] The parameter upload module 30 is used to take the trained local summary extraction model as the local model parameters for the current iteration round and send the local model parameters to a server belonging to the same federated learning system. The server receives the local model parameters and aggregates them with the local model parameters for the current iteration round sent by other client devices to obtain the global summary extraction model parameters for the current iteration round. The server then distributes the global summary extraction model parameters to each client device in the federated learning system as the local summary extraction model to be trained by each client device in the next iteration round.

[0146] From a software perspective, this application also provides an apparatus for performing all or part of the digest extraction method for multiple public security documents, see [link to relevant documentation]. Figure 7 The summary extraction device for multiple documents related to public safety specifically includes the following components:

[0147] The document receiving module 40 is used to receive a digest extraction request and obtain multiple target documents containing public safety content specified in the digest extraction request.

[0148] The multi-document heterogeneous graph generation module 50 is used to generate a multi-document heterogeneous graph corresponding to each of the target documents. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association relationship between the two nodes connected by the edge.

[0149] The topology graph partitioning module 60 is used to split the multi-document heterogeneous graph into separate topology graphs, each centered on a sentence node.

[0150] The model extraction module 70 is used to input each of the topological graphs into a local summary extraction model so that the local summary extraction model outputs summary data of the multi-document heterogeneous graphs corresponding to each of the target documents. The local summary extraction model is pre-trained based on the aforementioned training method for summary extraction models of public safety multi-documents.

[0151] The embodiments of the public security multi-document digest extraction device provided in this application can be used to execute the processing flow of the public security multi-document digest extraction method embodiments described above. Its functions will not be repeated here, but can be referred to the detailed description of the public security multi-document digest extraction method embodiments described above.

[0152] The digest extraction portion of the public security multi-document digest extraction device can be performed on a client device or a server. The choice can be made based on the processing capabilities of the client device and the limitations of the user's usage scenario. This application does not impose any limitations in this regard. If all operations are performed on the client device, the client device may further include a processor for the specific processing of the public security multi-document digest extraction.

[0153] This application also provides a federated learning system, specifically comprising: a server, and various client devices that are respectively connected to the server in communication;

[0154] The client device is used to execute the aforementioned method for training a summary extraction model for multiple public safety documents, and / or to execute the aforementioned method for extracting summaries for multiple public safety documents;

[0155] The client device is also used to send local model parameters in each iteration round to the server and receive global summary extraction model parameters sent by the server, so as to use the global summary extraction model parameters as the local summary extraction model to be trained in the next iteration round.

[0156] The server is configured to receive local model parameters sent by each of the client devices in each iteration round, aggregate the local model parameters of the current iteration round to obtain the global summary extraction model parameters of the current iteration round, and then send the global summary extraction model parameters to each of the client devices.

[0157] This application also provides an electronic device, which may include a processor, a memory, a receiver, and a transmitter. The processor is used to execute the summary extraction model training method for multiple public security documents and / or the summary extraction method for multiple public security documents mentioned in the above embodiments. The processor and the memory can be connected via a bus or other means, taking a bus connection as an example. The receiver can be connected to the processor and the memory via wired or wireless means.

[0158] The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0159] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the public security multi-document summary extraction model training method and / or the public security multi-document summary extraction method in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the public security multi-document summary extraction model training method and / or the public security multi-document summary extraction method in the above method embodiments.

[0160] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0161] The one or more modules are stored in the memory, and when executed by the processor, they execute the summary extraction model training method for public safety multiple documents and / or the summary extraction method for public safety multiple documents in the embodiment.

[0162] In some embodiments of this application, the user equipment may include a processor, a memory, and a transceiver unit. The transceiver unit may include a receiver and a transmitter. The processor, memory, receiver, and transmitter may be connected via a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0163] As one implementation method, the functions of the receiver and transmitter in this application can be implemented by transceiver circuits or dedicated transceiver chips, and the processor can be implemented by dedicated processing chips, processing circuits or general-purpose chips.

[0164] As another implementation approach, the server provided in this application embodiment can be implemented using a general-purpose computer. That is, the program code implementing the processor, receiver, and transmitter functions is stored in memory, and the general-purpose processor implements the processor, receiver, and transmitter functions by executing the code in memory.

[0165] This application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of the aforementioned method for training a summary extraction model for multiple public security documents and / or the method for summarizing multiple public security documents. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0166] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave.

[0167] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0168] In this application, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0169] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to the embodiments of this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for training a summary extraction model for multiple documents in public safety, characterized in that, include: A multi-document heterogeneous graph is generated based on multiple original documents containing public safety content. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association between the two nodes connected by the edge. The local summary extraction model, consisting of a graph convolutional network, a graph attention network, and a reinforcement learning model connected sequentially, is trained using the multi-document heterogeneous graph. This local summary extraction model is used to output multiple target sentences corresponding to the multi-document heterogeneous graph based on the input multi-document heterogeneous graph, thereby forming summary data of the multi-document heterogeneous graph. The process involves training a local summarization extraction model using the multi-document heterogeneous graph, which consists of a graph convolutional network, a graph attention network, and a reinforcement learning model connected sequentially. This local summarization extraction model is used to output multiple target sentences corresponding to the input multi-document heterogeneous graph, forming summary data for the multi-document heterogeneous graph. This includes: The multi-document heterogeneous graph is split into separate topological graphs, each centered on one of the sentence nodes. A local summarization extraction model is trained based on each of the aforementioned topologies, consisting of sequentially connected graph convolutional networks, graph attention networks, and reinforcement learning models. The graph convolutional networks update the node representations of each input topology, outputting updated topologies for each topology. The graph attention networks then extract cross-document information from each sentence node in the updated topologies and update the sentence representations of each sentence node accordingly, outputting the sentence representations corresponding to the multi-document heterogeneous graph. Finally, the reinforcement learning model, based on a pre-defined multi-step scenario Markov decision process strategy, extracts multiple target sentences corresponding to the multi-document heterogeneous graph through multiple time steps based on the input sentence representations, forming the summary data for the multi-document heterogeneous graph. After training the local summarization extraction model, which consists of a graph convolutional network, a graph attention network, and a reinforcement learning model connected sequentially, using the multi-document heterogeneous graph, the method further includes: The trained local summarization extraction model is used as the local model parameters for the current iteration round. These local model parameters are then sent to a server belonging to the same federated learning system. The server receives these local model parameters and aggregates them with the local model parameters for the current iteration round sent by other client devices to obtain the global summarization extraction model parameters for the current iteration round. The server then distributes these global summarization extraction model parameters to each client device in the federated learning system as the local summarization extraction model to be trained by each client device in the next iteration round.

2. The method for training a summary extraction model for multiple documents in public safety according to claim 1, characterized in that, The non-sentence nodes include: document nodes, word nodes, and entity nodes; each document node is used to represent a different document; each sentence node is used to represent a sentence in a document in a one-to-one manner; each word node is used to represent a word in a sentence in a one-to-one manner; and each entity node is used to represent a public safety entity in the document in a one-to-one manner. Correspondingly, the association types of each node in the multi-document heterogeneous graph include: edges between different document nodes, edges between a document node and a sentence node, edges between a sentence node and a word node, edges between a sentence node and an entity node, edges between different sentence nodes, edges between different word nodes, and edges between different entity nodes.

3. The method for training a summary extraction model for multiple documents in public safety according to claim 2, characterized in that, The process of generating a corresponding multi-document heterogeneous graph based on multiple original documents containing public safety content includes: Each document containing public safety content is used as a different document node. Each sentence is extracted from each document and each sentence is used as a different sentence node. Each word is extracted from each sentence and each word is used as a different word node. Furthermore, each public safety entity term is extracted from each document and each public safety entity term is used as a different entity node. Based on the document nodes, sentence nodes, word nodes, and entity nodes, as well as the relationships between different nodes, a multi-document heterogeneous graph corresponding to each document containing public safety content is constructed.

4. The method for training a summary extraction model for multiple documents in public safety according to claim 1, characterized in that, The reinforcement learning model is implemented based on an agent, which includes a global sentence encoder, a history extraction encoder, and an extractor connected in sequence. The global sentence encoder is used to generate a global context information representation corresponding to each sentence representation for each sentence representation corresponding to the multi-document heterogeneous graph output by the graph attention network, and then send each sentence representation and the global context information representation corresponding to each sentence representation to the extraction history encoder and the extractor respectively. The extraction history encoder is used to generate, in each time step, a historical information representation corresponding to each of the remaining sentences in the multi-document heterogeneous graph of the current time step (excluding the extracted target sentence), based on the sentence representation and the global context information representation corresponding to each of the remaining sentences, and send each of the historical information representations to the extractor. The extractor is used to aggregate the sentence representation, global context information representation, and historical information representation corresponding to each of the remaining sentences in each time step to obtain the connection representation corresponding to each of the remaining sentences. Then, it calculates the score corresponding to each of the remaining sentences based on the connection representation corresponding to each of the remaining sentences, and obtains the summary extraction probability corresponding to the current time step based on a preset multi-head pooling layer. Then, it determines whether the summary extraction probability is greater than a preset probability threshold. If not, the remaining sentence with the highest score in the current time step is extracted and used as the extracted target sentence in the next time step. If so, the execution of the next time step is stopped, and each of the target sentences in the current time step is used as the summary data of the multi-document heterogeneous graph, and the summary data is output.

5. A method for extracting summaries from multiple documents for public safety, characterized in that, Also includes: Receive a digest extraction request and retrieve multiple target documents containing public safety content specified in the digest extraction request; Generate a multi-document heterogeneous graph corresponding to each of the target documents. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association relationship between the two nodes connected by the edge. The multi-document heterogeneous graph is split into separate topological graphs, each centered on one of the sentence nodes. Each of the topological graphs is input into a local summary extraction model, so that the local summary extraction model outputs summary data of the multi-document heterogeneous graph corresponding to each of the target documents, wherein the local summary extraction model is pre-trained based on the summary extraction model training method for public safety multi-documents as described in any one of claims 1 to 4.

6. A training device for a multi-document summary extraction model for public safety, characterized in that, include: The heterogeneous graph generation module is used to generate a corresponding multi-document heterogeneous graph based on multiple original documents containing public safety content. The semantic types of each node in the multi-document heterogeneous graph are of various types, including sentence nodes and non-sentence nodes. The edges between the nodes in the multi-document heterogeneous graph are used to represent the association relationship between the two nodes connected by the edge. The local model training module is used to train a local summary extraction model consisting of a graph convolutional network, a graph attention network, and a reinforcement learning model connected in sequence using the multi-document heterogeneous graph, so that the local summary extraction model can output multiple target sentences corresponding to the multi-document heterogeneous graph based on the input multi-document heterogeneous graph, so as to form the summary data of the multi-document heterogeneous graph. The step of training a local summarization extraction model using the multi-document heterogeneous graph (MHP) consists of a graph convolutional network, a graph attention network, and a reinforcement learning model connected sequentially. This MHP is used to output multiple target sentences corresponding to the input MHP to form summary data of the MHP, including: The multi-document heterogeneous graph is split into separate topological graphs, each centered on one of the sentence nodes. A local summarization extraction model is trained based on each of the aforementioned topologies, consisting of sequentially connected graph convolutional networks, graph attention networks, and reinforcement learning models. The graph convolutional networks update the node representations of each input topology, outputting updated topologies for each topology. The graph attention networks then extract cross-document information from each sentence node in the updated topologies and update the sentence representations of each sentence node accordingly, outputting the sentence representations corresponding to the multi-document heterogeneous graph. Finally, the reinforcement learning model, based on a pre-defined multi-step scenario Markov decision process strategy, extracts multiple target sentences corresponding to the multi-document heterogeneous graph through multiple time steps based on the input sentence representations, forming the summary data for the multi-document heterogeneous graph. After training the local summarization extraction model, which consists of a graph convolutional network, a graph attention network, and a reinforcement learning model, using the multi-document heterogeneous graph, the summarization extraction model training device for public safety multi-documents is further configured to perform: The trained local summarization extraction model is used as the local model parameters for the current iteration round. These local model parameters are then sent to a server belonging to the same federated learning system. The server receives these local model parameters and aggregates them with the local model parameters for the current iteration round sent by other client devices to obtain the global summarization extraction model parameters for the current iteration round. The server then distributes these global summarization extraction model parameters to each client device in the federated learning system as the local summarization extraction model to be trained by each client device in the next iteration round.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for training a summary extraction model for multiple documents for public safety as described in any one of claims 1 to 4, and / or implements the method for summarizing multiple documents for public safety as described in claim 5.

8. A federated learning system, characterized in that, include: The server, and the various client devices that communicate with the server respectively; The client device is used to execute the summary extraction model training method for multiple documents in public safety as described in any one of claims 1 to 4, and / or to execute the summary extraction method for multiple documents in public safety as described in claim 5; The client device is also used to send local model parameters in each iteration round to the server and receive global summary extraction model parameters sent by the server, so as to use the global summary extraction model parameters as the local summary extraction model to be trained in the next iteration round. The server is configured to receive local model parameters sent by each of the client devices in each iteration round, aggregate the local model parameters of the current iteration round to obtain the global summary extraction model parameters of the current iteration round, and then send the global summary extraction model parameters to each of the client devices.

Images