Document-level event extraction model training method, event extraction method and system

Abstract

The invention relates to the technical field of document-level event extraction, and provides a document-level event extraction model training method and an event extraction method and system in order to solve the problems that in the prior art, remote dependence modeling capacity is insufficient, noise interference is likely to happen, and event output semantics are inconsistent. The document-level event extraction model training method comprises the following steps: preprocessing a document sample and initializing nodes; constructing an anaphora enhanced heterogeneous graph; performing two-stage graph information aggregation; generating global event classification task loss; generating event decoding task loss; summing the loss of the event decoding task and the loss of the weighted global event classification task, taking the sum as the total loss of the document-level event extraction model, and jointly optimizing the parameters of the document-level event extraction model through multi-task learning. According to the method, not only can the fingering be accurately utilized as an explicit semantic bridge to strengthen remote reasoning, but also associated noise can be effectively filtered, and meanwhile, document-level global semantic constraints can be fused to ensure output consistency.

Description

Technical Field

[0001] This invention relates to the field of document-level event extraction technology, and in particular to a document-level event extraction model training method, event extraction method, and system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Document-level event extraction is one of the core tasks in the field of information extraction. Its goal is to extract complete and structured event records from unstructured whole texts. Unlike sentence-level tasks, real-world documents (such as financial announcements and legal documents) generally exhibit an "argument dispersion" phenomenon: multiple arguments of an event (such as subject, object, time, and location) are often scattered across different paragraphs or sentences of the document, and a single document often contains multiple related event instances. This characteristic requires the model to have strong long-distance semantic association and reasoning capabilities.

[0004] The groundbreaking work Doc2EDAG formalized the event table filling task as a sequence generation problem of a directed acyclic graph, providing a new approach to end-to-end multi-event extraction. Subsequently, models represented by GIT (Graph-based Interaction Model with a Tracker) employed heterogeneous graph neural networks to aggregate document-level contextual information by constructing an interaction graph of sentences and entity mentions, further improving performance. However, these methods typically model documents as static sets of entity mentions, with information flow heavily reliant on fixed graph structures based on sentence order or adjacency relationships. When key entities (such as company names) appear only at the beginning of the document, and are referred to extensively by pronouns (such as "the company," "its"), this implicit, co-occurrence-based reasoning pattern struggles to establish effective long-range connections, resulting in low recall for scattered arguments.

[0005] Existing document-level relation extraction methods demonstrate the effectiveness of explicitly modeling referential information in improving cross-sentence reasoning by explicitly introducing referential nodes and assigning appropriate weights to them using attention mechanisms. However, directly transferring the referential modeling strategy from relation extraction to event extraction is not feasible. This is because event extraction involves denser argument role assignments and more complex entity interactions; a simple 'fully connected' strategy would introduce a large amount of spurious referential noise into the graph, leading to overly smooth node representations. The consequence is that the model struggles to distinguish entity clusters belonging to different events, impairing the clarity of event boundaries and the accuracy of argument role discrimination. Furthermore, current mainstream event decoders often employ bottom-up local combination strategies, lacking explicit awareness of the document's macro-theme or intent, and are prone to generating pseudo-event records that conflict with the semantic context of the entire text. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a document-level event extraction model training method, event extraction method, and system. This method can accurately utilize pronouns as explicit semantic bridges to enhance remote reasoning, effectively filter associated noise, and incorporate document-level global semantic constraints to ensure output consistency.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for training a document-level event extraction model.

[0008] In one or more embodiments, a document-level event extraction model training method includes: Extract sentences, entity mentions, and pronouns exceeding a set frequency threshold from document samples, initialize them as sentence nodes, entity mention nodes, and pronoun nodes in a heterogeneous graph, and obtain the corresponding initial feature representations; Using the initialized nodes and their initial feature representations, a heterogeneous graph with enhanced referentiality is constructed, which includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges, and referential-entity soft connection edges. Information from referential nodes is aggregated by soft linking referential nodes and noise is filtered by gating to enhance the feature representation of entity mention nodes. Then, together with the feature representation of sentence nodes, multi-layer message passing is performed in the heterogeneous graph with enhanced referentiality to complete global structural reasoning and update the feature representation of all nodes. Pooling is performed on the feature representations of all sentence nodes to obtain a document-level global vector. The event type distribution of document samples is predicted by a multi-label classifier to generate a global event classification task loss. Based on the updated feature representations of all nodes, a tracker-based decoder is used to progressively predict event types and corresponding arguments to obtain structured event records and generate event decoding task loss. The loss from the event decoding task is summed with the weighted loss from the global event classification task to form the total loss of the document-level event extraction model. Then, the parameters of the document-level event extraction model are jointly optimized through multi-task learning.

[0009] As one implementation, the referential-entity soft link edge is: dynamically calculating the entity mention node through a gating attention mechanism. With reference node Association weight between This forms a virtual semantic edge; its calculation formula is: ; ; in, The feature dimension of the key vector; To query the transformation matrix, used to mention the entity node. The feature representation is mapped to the query space; For the first The feature representation of a node mentioned by an entity; This is a transpose operation for a matrix or vector. This is the key transformation matrix, used to transform the reference nodes. The feature representation is mapped to the key space; For the first The feature representation of each node; To refer to node dimension The normalized exponential function on; Nodes mentioned for entities With reference node The degree of semantic matching between them.

[0010] As one implementation method, in the process of enhancing the feature representation of entity mention nodes, the entity mention nodes aggregate the information of the referent nodes through referent-entity soft link edges, introduce gated filtering of noise, and obtain the enhanced feature representation of the entity mention nodes: ; ; in, Feature representation of the enhanced entity mention node; Presentation layer normalization operation; For the first The feature representation of a node mentioned by an entity; This refers to the set of all referential nodes in the heterogeneous graph that is being referred to; This is the value transformation matrix, used to map the feature representation of the referent node to the semantic value space; For the first The feature representation of each node; Nodes mentioned for entities With reference node The correlation weight between them; The gating factor; Use the Sigmoid activation function; To make the first The feature representation of the entity mentioning the node and the first entity mentioning the node are related to the feature representation of the node. The feature representations of each node are concatenated; This represents the bias feature.

[0011] As one implementation method, a pre-trained language model is used to obtain initial feature representations of sentence nodes, entity mention nodes, and referential nodes.

[0012] As one implementation method, the expression for the total loss of the document-level event extraction model is: ; ; ; in, Total loss; Loss due to event decoding task; Classify task losses for global events; To balance the hyperparameters; A collection of actual event records in the document; For document-based Predict the type of event generated and its set of arguments The conditional probability; The total number of predefined event types; To indicate whether a document actually contains the first... The label for the event class is 1 for inclusion and 0 for exclusion; For predicting documents containing the first The probability of a class of events.

[0013] A second aspect of the present invention provides a document-level event extraction model training system.

[0014] In one or more embodiments, a document-level event extraction model training system includes: The document preprocessing and node initialization module is used to extract sentences, entity mentions and pronouns exceeding a set frequency threshold from document samples, and initialize them as sentence nodes, entity mention nodes and pronoun nodes of heterogeneous graphs and obtain the corresponding initial feature representations. The module for constructing heterogeneous graphs with enhanced referentiality is used to construct heterogeneous graphs with enhanced referentiality using initialized nodes and their initial feature representations. It includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges, and referential-entity soft connection edges. The two-stage graph information aggregation module is used to aggregate the information of reference nodes through reference-entity soft connection edges and introduce gated filtering noise to enhance the feature representation of entity mention nodes. Then, together with the feature representation of sentence nodes, it performs multi-layer message passing in the reference-enhanced heterogeneous graph to complete global structural reasoning and update the feature representation of all nodes. The global event classification task loss generation module is used to pool the feature representations of all sentence nodes to obtain a document-level global vector, predict the event type distribution of document samples through a multi-label classifier, and generate the global event classification task loss. The event decoding task loss generation module is used to predict the event type and corresponding arguments step by step based on the updated feature representation of all nodes using a tracker-based decoder, to obtain a structured event record and generate the event decoding task loss. The total loss construction and parameter optimization module is used to sum the event decoding task loss and the weighted global event classification task loss as the total loss of the document-level event extraction model, and then jointly optimize the parameters of the document-level event extraction model through multi-task learning.

[0015] A third aspect of the present invention provides an event extraction method.

[0016] In one or more embodiments, an event extraction method includes: Obtain the document to be processed, which is in the same domain as the document sample used to train the document-level event extraction model; A document-level event extraction model is used to obtain structured event records from the document to be processed; wherein the document-level event extraction model is trained using the steps in the document-level event extraction model training method described above. The document-level event extraction model is configured as follows: Extract sentences, entity mentions, and pronouns exceeding a set frequency threshold from the document to be processed, initialize them as sentence nodes, entity mention nodes, and pronoun nodes in a heterogeneous graph, and obtain the corresponding initial feature representations; Using the initialized nodes and their initial feature representations, a heterogeneous graph with enhanced referentiality is constructed, which includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges, and referential-entity soft connection edges. Information from referential nodes is aggregated by soft linking referential nodes and noise is filtered by gating to enhance the feature representation of entity mention nodes. Then, together with the feature representation of sentence nodes, multi-layer message passing is performed in the heterogeneous graph with enhanced referentiality to complete global structural reasoning and update the feature representation of all nodes. Pooling is performed on the feature representations of all sentence nodes to obtain a document-level global vector, which is then used to predict the event type distribution of document samples by a multi-label classifier; Based on the updated feature representations of all nodes, a tracker-based decoder is used to progressively predict event types and corresponding arguments to obtain structured event records.

[0017] A fourth aspect of the present invention provides an event extraction system.

[0018] In one or more embodiments, an event extraction system includes: The document acquisition module is used to acquire documents to be processed, which are in the same domain as the document samples used to train the document-level event extraction model. An event extraction module is used to obtain structured event records from the document to be processed using a document-level event extraction model; wherein the document-level event extraction model is trained using the steps in the document-level event extraction model training method described above. The document-level event extraction model is configured as follows: Extract sentences, entity mentions, and pronouns exceeding a set frequency threshold from the document to be processed, initialize them as sentence nodes, entity mention nodes, and pronoun nodes in a heterogeneous graph, and obtain the corresponding initial feature representations; Using the initialized nodes and their initial feature representations, a heterogeneous graph with enhanced referentiality is constructed, which includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges, and referential-entity soft connection edges. Information from referential nodes is aggregated by soft linking referential nodes and noise is filtered by gating to enhance the feature representation of entity mention nodes. Then, together with the feature representation of sentence nodes, multi-layer message passing is performed in the heterogeneous graph with enhanced referentiality to complete global structural reasoning and update the feature representation of all nodes. Pooling is performed on the feature representations of all sentence nodes to obtain a document-level global vector, which is then used to predict the event type distribution of document samples by a multi-label classifier; Based on the updated feature representations of all nodes, a tracker-based decoder is used to progressively predict event types and corresponding arguments to obtain structured event records.

[0019] A fifth aspect of the present invention provides a computer-readable storage medium.

[0020] In one or more embodiments, a computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps in the document-level event extraction model training method described above.

[0021] In other embodiments, a computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the event extraction method as described above.

[0022] A sixth aspect of the present invention provides an electronic device.

[0023] In one or more embodiments, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the document-level event extraction model training method as described above.

[0024] In other embodiments, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the event extraction method described above.

[0025] Compared with the prior art, the beneficial effects of the present invention are: (1) In the process of training the document-level event extraction model, this invention constructs a heterogeneous graph with enhanced referentiality, aggregates the information of referential nodes based on the soft connection edges between referentiality and entity, and introduces gated filtering to enhance the feature representation of entity mention nodes. This enables explicit screening and establishment of high-confidence referential associations. The enhanced feature representation of entity mention nodes and the feature representation of sentence nodes are passed through multiple layers of message passing to complete global structural reasoning and update the feature representation of all nodes, thereby constructing a cross-sentence semantic bridge and suppressing the propagation of irrelevant noise. Then, the total loss of the document-level event extraction model is calculated by using the event decoding task loss and the weighted global event classification task loss. This realizes the comprehensive global event constraints and injects semantic supervision into the document level with the help of a multi-task learning mechanism, thereby improving the internal consistency of event records.

[0026] (2) This invention utilizes a pre-trained document-level event extraction model to extract document-level events from documents in the same domain as the document samples used to train the model. During the document-level event extraction process, a heterogeneous graph with enhanced referentiality is constructed by introducing referential nodes and gated attention. This effectively solves the problem of long-range reasoning breakage caused by the implicit referential chain in long documents. It can accurately utilize referential words as explicit semantic bridges to strengthen long-range reasoning and effectively filter associated noise. At the same time, it achieves internal consistency of event records. Attached Figure Description

[0027] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0028] Figure 1 This is a flowchart of a document-level event extraction model training method according to an embodiment of the present invention; Figure 2 This is the overall architecture of the document-level event extraction model in this embodiment of the invention; Figure 3 This is a systematic experimental comparison of the document-level event extraction model of this invention and the traditional GIT model on the public benchmark dataset ChFinAnn; Figure 4 This is a comparison between the document-level event extraction model of this invention and the baseline model under different span parameters; Figure 5 This is an example of a gated attention mechanism in an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a document-level event extraction model training system according to an embodiment of the present invention; Figure 7 This is a flowchart of an event extraction method according to an embodiment of the present invention; Figure 8 This is a schematic diagram of an event extraction system according to an embodiment of the present invention; Figure 9 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in the embodiments of the invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0031] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0032] Graph neural networks (Graph Neural Networks) are widely used for document-level information extraction due to their powerful structural modeling capabilities. In the field of relation extraction, the LSR (Latent Space Representation Model) uses document structure as a latent variable for inference, while the GAIN model (which constructs a two-layer graph structure of mention-level graph (hMG) and entity-level graph (EG) to fuse heterogeneous information in documents) builds a two-layer graph of mentions and entities to fuse heterogeneous information. Inspired by this, the GIT model was the first to introduce heterogeneous graph interaction into document-level event extraction, building a fully connected graph through sentence nodes and mention nodes to propagate global information, achieving significant progress. However, its graph construction is based on fixed physical or fully connected rules, failing to fully utilize explicit linguistic cues such as pronouns to construct more accurate semantic paths.

[0033] Dereference resolution is a core task in connecting scattered references in text. Existing technologies innovatively treat pronouns as nodes in a graph and enhance relation reasoning by connecting all potential reference pairs, demonstrating the effectiveness of explicit reference modeling. However, event extraction is more complex than relation extraction, requiring the identification of multiple arguments with specific roles from a large number of candidate entities. In event documents with dense arguments, dense connection strategies introduce significant noise, leading to overly smooth graph node representations. To address this issue, this invention introduces a learnable gating mechanism to dynamically filter low-quality reference associations while preserving high-confidence semantic bridges.

[0034] Document-level event extraction aims to extract structured event records from complete documents. Its challenge lies in "argument dispersion," where arguments for a single event often span multiple widely spaced sentences, placing high demands on the model's cross-sentence reasoning capabilities. Existing heterogeneous graph-based models mostly rely on the physical proximity between sentences for implicit information transfer, failing to effectively utilize explicit semantic associations represented by pronouns. This results in insufficient long-range dependency modeling capabilities and susceptibility to noise interference. Simultaneously, mainstream bottom-up decoding mechanisms lack explicit constraints on the overall document theme, easily leading to semantically inconsistent event outputs. To address these issues, this invention constructs a pronoun-enhanced heterogeneous graph structure (GA-GIT, GraphAttention Graph-based Interaction Model with a Tracker architecture). Through a designed gating attention mechanism, it can explicitly filter and establish high-confidence pronoun associations, thereby building cross-sentence semantic bridges and suppressing the propagation of irrelevant noise. Furthermore, a global event constraint module is introduced, injecting document-level semantic supervision through a multi-task learning mechanism to improve the internal consistency of event records. Experiments on the financial dataset ChFinAnn show that the GA-GIT architecture of this invention achieves a Micro-F1 score of 82.2%, which is significantly better than existing mainstream baseline models. The improvement in recall is particularly significant in long-distance extraction scenarios where arguments are highly dispersed.

[0035] Figure 1 A schematic diagram illustrating the principle of the document-level event extraction model training method according to an embodiment of the present invention is provided. Combined with... Figure 1 and Figure 2 The document-level event extraction model training method of this embodiment may include the following steps S101 to S106.

[0036] The specific implementation process of steps S101 to S106 is as follows: Step S101: Extract sentences, entity mentions, and pronouns exceeding a set frequency threshold from the document samples, initialize them as sentence nodes, entity mention nodes, and pronoun nodes in a heterogeneous graph, and obtain the corresponding initial feature representations.

[0037] Given a containing Document of 1 sentence The goal of document-level event extraction is to extract a set of event records. Each event is logged. By a predefined event type and its corresponding set of arguments Composition, in which Indicates the type of argument role. Entity mentions in the document. In addition, to explicitly model cross-sentence anaphora, embodiments of the present invention define a set of anaphora candidate items extracted from the document , as a potential semantic hub for connecting scattered arguments; is the total number of extracted pronouns; is the set of event types.

[0038] In the specific implementation process, the document sample is segmented into sentences and entities are recognized, and pronouns exceeding the set frequency threshold in the document are extracted; each sentence, entity mention, and pronoun are respectively initialized as sentence nodes, entity mention nodes, and anaphora nodes in the heterogeneous graph, and the initial feature representations of each node are obtained using a pre-trained language model.

[0039] In some alternative embodiments, existing open-source tools, such as spaCy, can be used to segment the document sample into sentences and recognize entities, and extract pronouns exceeding the set frequency threshold in the document.

[0040] For example, directly call the NLP pipeline of the spaCy toolkit; use the sentencizer component of spaCy to segment the document and segment the document sample into sentences; Use the Part-of-Speech Tagging (POS Tagging) function of spaCy to identify words with a POS tag of PRON (pronoun) and text in a preset high-frequency word list (such as "its", "the") as anaphora nodes.

[0041] Use the built-in NER component of spaCy (or load a fine-tuned model) to obtain the boundaries of entity mentions and implement entity recognition of the document sample.

[0042] The document is expanded into a Token sequence , and input into a domain-adapted pre-trained language model (such as the Mengzi-BERT-base-fi model) to obtain context embeddings , this matrix contains all Tokens in the document -dimensional feature representations. Among them, represents the length of the Token sequence input into the pre-trained language model, that is, the total number of Tokens contained in the document after tokenization; represents the feature dimension (or hidden layer dimension) output by the pre-trained language model, that is, the number of values of the semantic feature vector used by the model to represent each Token.

[0043] The initial feature representations of sentence nodes and entity mention nodes are obtained through max pooling operations respectively: ; ; in, The dimension of the output of the pre-trained language model is The full document context embedding matrix; and The first The start and end index positions of each sentence node in the document sequence; and The first Each entity mentions the start and end index positions of the node in the document sequence; This is a max pooling operation along the sequence length dimension, used to aggregate a variable-length sequence of token vectors into a feature vector of fixed dimensions; and The first and second parts obtained after aggregation are respectively The sentence node and the first The initial feature vectors of the entity mentions nodes, each with dimension 1. .

[0044] Based on sentence nodes and entity mention node The initial feature representation yields the sentence node feature set. and entity node feature set .

[0045] Step S102: Using the initialized nodes and their initial feature representations, construct a heterogeneous graph with enhanced referentiality, which includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges, and referential-entity soft connection edges.

[0046] The heterogeneous graph G with enhanced representation contains three types of nodes: sentence nodes. Entity mention node and referencing nodes The graph contains four types of edges: (1) Sentence-Sentence Edge : Based on sentence adjacency or discourse structure connections; (2) Sentence-entity edge : Connects a sentence to its contained entity references; (3) Entity-Entity Edge : Connects entities within the same sentence or those with explicit coreference; (4) Referent - Entity soft link edge Dynamically calculate entity mention nodes through a gating attention mechanism. With reference node Association weight between This forms a virtual semantic edge; its calculation formula is: ; ; in, The feature dimension of the key vector is used to adjust the value of the inner product to prevent the Softmax function from entering the saturation region where the gradient is minimal due to excessively large values. To query the transformation matrix, used to mention the entity node. The feature representation is mapped to the query space; For the first The feature representation of a node mentioned by an entity; This is a transpose operation for a matrix or vector. This is the key transformation matrix, used to transform the reference nodes. The feature representation is mapped to the key space; For the first The feature representation of each node; To refer to node dimension The normalized exponential function on; Nodes mentioned for entities With reference node The degree of semantic matching between them. The larger the value, the more likely it is that the node is considered to be referred to. Referencing entity mention node The higher the probability, the better.

[0047] Step S103: Aggregate the information of the reference nodes through the soft link edges of reference and entity and introduce gating to filter noise to enhance the feature representation of the entity mention nodes. Then, together with the feature representation of the sentence nodes, perform multi-layer message passing in the heterogeneous graph of reference enhancement to complete global structural reasoning and update the feature representation of all nodes.

[0048] In the process of enhancing the feature representation of entity mention nodes, entity mention nodes aggregate information from reference nodes through reference-entity soft connection edges, and introduce gated filtering to filter noise, resulting in the enhanced feature representation of entity mention nodes: ; ; in, Feature representation of the enhanced entity mention node; Presentation layer normalization operation; For the first The feature representation of a node mentioned by an entity; This refers to the set of all referential nodes in the heterogeneous graph that is being referred to; This is the value transformation matrix, used to map the feature representation of the referent node to the semantic value space; For the first The feature representation of each node; Nodes mentioned for entities With reference node The correlation weight between them; The gating coefficients, calculated by the Sigmoid activation function, range from (0,1) and are used to explicitly filter noise; when When the value approaches 0, it indicates that the model judges the connection as noise and blocks it. Use the Sigmoid activation function; To make the first The feature representation of the entity mentioning the node and the first entity mentioning the node are related to the feature representation of the node. The feature representations of each node are concatenated; This represents the bias feature.

[0049] Entity reference nodes aggregate referential node information through gated soft links, achieving cross-sentence semantic enhancement. The enhanced entity nodes and sentence nodes are then linked in sentence-to-sentence edges within the heterogeneous graph. Sentence-Entity Edge and entity-entity edge Multiple rounds of message passing are performed to complete global structural reasoning and update the features of all nodes.

[0050] Step S105: Pool the feature representations of all sentence nodes to obtain a document-level global vector. Through multiple methods, obtain structured event records and generate the event decoding task loss.

[0051] To improve the global consistency of event type prediction, this embodiment of the invention adds an auxiliary task of document-level event type classification. The global document vector is obtained by max pooling all sentence representations. Then, a classifier is used to predict the event type distribution of the document: ; The binary cross-entropy loss for this task As a global constraint signal.

[0052] in, This represents the predicted document-level event type probability distribution vector; This represents the Sigmoid activation function; This represents the weight matrix of the classifier; This represents the bias vector of the classifier.

[0053] Step S106: Sum the event decoding task loss and the weighted global event classification task loss as the total loss of the document-level event extraction model, and then jointly optimize the parameters of the document-level event extraction model through multi-task learning.

[0054] The decoder employs a tracker-based sequence generation strategy to progressively generate event types and arguments, with its loss denoted as... .

[0055] The expression for the total loss of the document-level event extraction model is: ; ; ; in, Total loss; Loss due to event decoding task; Classify task losses for global events; To balance the hyperparameters, used to adjust the contribution of the global classification task to the total loss; A collection of actual event records in the document; For document-based Predict the type of event generated and its set of arguments The conditional probability; The total number of predefined event types; To indicate whether a document actually contains the first... The label for the event class is 1 for inclusion and 0 for exclusion; For predicting documents containing the first The probability of a class of events.

[0056] Data sets and metrics: This invention evaluates the model's performance on the large-scale Chinese financial event extraction dataset ChFinAnn. This dataset consists of 32,040 real financial announcements, covering five common types of financial events (share pledge, share reduction, share increase, share freeze, and share repurchase). Compared to other datasets, ChFinAnn exhibits significant "argument scattering": the arguments for an event record are scattered across an average of six sentences, and 29% of documents contain multiple event records. This severely tests the model's ability to perform cross-sentence reasoning in long documents.

[0057] This invention adopts the standard partitioning method, with training, validation, and test sets of 2632, 3204, and 3204 documents respectively. The evaluation metrics used are standard micro-average precision (P), recall (R), and F1 score. Following previous work, this invention primarily focuses on the performance of the event record extraction task, which is the core metric for measuring the model's comprehensive structured extraction capability.

[0058] The Mengzi-BERT-base-fin encoder was used. The graph neural network had 3 layers, a gated attention dropout rate of 0.1, and a hyperparameter λ=0.05. The AdamW optimizer was used with a learning rate of 2e-5 and a batch size of 64. All experiments were performed on an NVIDIA RTX 3090 GPU, and the reported results are the average of 3 independent runs.

[0059] This invention compares GA-GIT with the following mainstream document-level event extraction models: DCFEE: An early pipeline model that completes arguments through center sentence detection and neighbor expansion strategies.

[0060] Doc2EDAG: An end-to-end sequence generation model based on Transformer, which models the extraction task as EDAG path generation.

[0061] GIT: The current state-of-the-art model, utilizing heterogeneous graph networks and a Tracker mechanism for extraction. GA-GIT retains its basic architecture but improves upon the graph interaction mechanism.

[0062] Table 1 shows the comparison results of different models on the test set.

[0063] Table 1 compares the performance of event record extraction on the test set;

[0064] As can be observed from Table 1, in terms of overall performance improvement, the GA-GIT of this embodiment of the invention has achieved the latest optimal performance, with a Micro-F1 value of 82.2%, which is 1.9 percentage points higher than the baseline and GIT, thus verifying the effectiveness of the GA-GIT referential enhancement strategy of this embodiment of the invention.

[0065] Significantly Improved Recall: Notably, compared to the improvement in precision (+0.5%), GA-GIT shows a more significant improvement in recall (+3.4%), supporting the core assumption of this invention: existing graph-based methods (such as GIT) often miss arguments due to broken implicit connections. In contrast, the referential enhancement heterogeneous graph of this invention establishes explicit semantic bridges through pronouns, successfully recovering these "hard-to-find" arguments (especially the main entities referenced by pronouns), thereby significantly improving recall without sacrificing precision.

[0066] To investigate the contribution of each component in the GA-GIT of this embodiment, an ablation experiment was conducted by removing a specific module, and the results are reported in Table 2.

[0067] Table 2. Ablation experiment analysis;

[0068] The experimental analysis is as follows: The necessity of Anaphor Nodes: Removing Anaphor Nodes (without Anaphor Nodes) resulted in the largest performance drop (-1.7%), proving that explicit modeling of Anaphor information is crucial for capturing cross-sentence dependencies, and that relying solely on physical adjacency edges cannot effectively transmit long-distance semantics.

[0069] The denoising effect of gating: If referential nodes are retained but gating is removed (without gating), the performance decreases by 0.9%. This confirms that in argument-dense event graphs, simple fully connected connections introduce a large amount of false coreference noise, while gating effectively acts as a "semantic filter," preventing the graph from becoming too smooth.

[0070] Corrective effect of global constraints: Removing the global constraint (without global constraint) resulted in a 0.6% decrease in F1 score. Further analysis revealed that this variant readily predicts mutually exclusive event types in multi-event documents, indicating that global semantic supervision helps regulate the decoder's behavior and ensures that local predictions are consistent with the document's main theme.

[0071] To further verify the model's ability to handle the "argument dispersion" problem, this embodiment of the invention follows the GIT settings, dividing the test set samples into four groups of increasing difficulty (sets I, II, III, IV) based on the number of sentences spanned by the event arguments. Among them, set I (within 1 sentence) represents the simple scenario with the most concentrated arguments, and set IV represents the extremely difficult scenario with the largest argument span (more than 10 sentences).

[0072] like Figure 4As shown, this embodiment of the invention compares the F1 performance of GA-GIT and GIT in different groups: it can be seen that the baseline model (red dashed line) shows a significant performance degradation in long-distance (set III and set IV) scenarios. In contrast, GA-GIT (blue solid line) exhibits stronger stability. In particular, in set IV where the distance is greater than 10, GA-GIT achieves a significant improvement of 4.0%, demonstrating its effectiveness in capturing long-distance contextual information.

[0073] In the simple set I, the performance difference between the two is small. However, as the argument dispersion increases (sets II and IV), the performance degradation trend of GIT becomes significant, indicating that implicit graph connections gradually fail over long distances. GA-GIT demonstrates strong robustness in the most challenging set IV, achieving a 4.0% F1 improvement compared to GIT. This trend intuitively proves that referential nodes, acting as "semantic relay stations," effectively shorten inference paths and solve the problem of long-distance dependency breaks.

[0074] To visually demonstrate the working mechanism of GA-GIT, this embodiment of the invention selects a specific test sample for visualization analysis, such as... Figure 5 As shown, the gating mechanism assigns high weights (g = 0.97) to the correct antecedents, thereby establishing a strong semantic bridge, while suppressing noisy links with irrelevant entities with low weights (g = 0.03).

[0075] In this case, the core entity "X Corporation" only appears in its full name in the first sentence of the document. The sixth sentence uses the pronoun "its" to refer to this entity and describes the "share reduction" action. The baseline model GIT failed to establish a connection between the first and sixth sentences, incorrectly omitting the argument of "shareholder." In contrast, GA-GIT successfully identified the pronoun "its" as a key node. Through a gated attention mechanism, the model established a high-weighted virtual edge between "its" and "X Corporation," creating a shortcut for semantic transmission and thus correctly extracting the complete event record. Figure 3 (a) Systematic experiments of the traditional GIT model on the public benchmark dataset ChFinAnn; Figure 3 (b) in the figure is the document-level event extraction model of the present invention, namely the systematic experiment of GA-GIT on the public benchmark dataset ChFinAnn; by comparison, it can be seen that the document-level event extraction model of the present invention can accurately utilize pronouns as explicit semantic bridges and enhance remote reasoning.

[0076] like Figure 6 As shown, the document-level event extraction model training system provided in this embodiment of the invention can be implemented in software. The document-level event extraction model training system includes the following software modules: The document preprocessing and node initialization module 601 is used to extract sentences, entity mentions and pronouns exceeding a set frequency threshold from the document samples, initialize them as sentence nodes, entity mention nodes and pronoun nodes of the heterogeneous graph and obtain the corresponding initial feature representations. The reference-enhanced heterogeneous graph construction module 602 is used to construct a reference-enhanced heterogeneous graph using initialized nodes and their initial feature representations. It includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges, and reference-entity soft connection edges. The two-stage graph information aggregation module 603 is used to aggregate the information of reference nodes through reference-entity soft connection edges and introduce gated filtering noise to enhance the feature representation of entity mention nodes. Then, together with the feature representation of sentence nodes, it performs multi-layer message passing in the reference-enhanced heterogeneous graph to complete global structural reasoning and update the feature representation of all nodes. The global event classification task loss generation module 604 is used to pool the feature representations of all sentence nodes to obtain a document-level global vector, predict the event type distribution of document samples through a multi-label classifier, and generate the global event classification task loss. The event decoding task loss generation module 605 is used to predict the event type and corresponding arguments step by step based on the updated feature representation of all nodes using a tracker-based decoder to obtain a structured event record and generate the event decoding task loss. The total loss construction and parameter optimization module 606 is used to sum the event decoding task loss and the weighted global event classification task loss as the total loss of the document-level event extraction model, and then jointly optimize the parameters of the document-level event extraction model through multi-task learning.

[0077] It should be noted that each module in the document-level event extraction model training system of this invention corresponds one-to-one with each step in the document-level event extraction model training method in the above embodiments, and their specific implementation processes are the same, so they will not be repeated here.

[0078] In one or more embodiments, such as Figure 7 As shown, an event extraction method includes: Step S701: Obtain the document to be processed, which is in the same domain as the document sample used to train the document-level event extraction model; Step S702: Using a document-level event extraction model, obtain structured event records from the document to be processed; wherein, the document-level event extraction model is trained using the steps in the document-level event extraction model training method described above; The document-level event extraction model is configured as follows: Extract sentences, entity mentions, and pronouns exceeding a set frequency threshold from the document to be processed, initialize them as sentence nodes, entity mention nodes, and pronoun nodes in a heterogeneous graph, and obtain the corresponding initial feature representations; Using the initialized nodes and their initial feature representations, a heterogeneous graph with enhanced referentiality is constructed, which includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges, and referential-entity soft connection edges. Information from referential nodes is aggregated by soft linking referential nodes and noise is filtered by gating to enhance the feature representation of entity mention nodes. Then, together with the feature representation of sentence nodes, multi-layer message passing is performed in the heterogeneous graph with enhanced referentiality to complete global structural reasoning and update the feature representation of all nodes. Pooling is performed on the feature representations of all sentence nodes to obtain a document-level global vector, which is then used to predict the event type distribution of document samples by a multi-label classifier; Based on the updated feature representations of all nodes, a tracker-based decoder is used to progressively predict event types and corresponding arguments to obtain structured event records.

[0079] In one or more embodiments, such as Figure 8 As shown, an event extraction system includes: The document acquisition module 801 is used to acquire the document to be processed, which is in the same domain as the document sample used to train the document-level event extraction model. The event extraction module 802 is used to obtain structured event records from the document to be processed using a document-level event extraction model; wherein the document-level event extraction model is trained using the steps in the document-level event extraction model training method described above. The document-level event extraction model is configured as follows: Extract sentences, entity mentions, and pronouns exceeding a set frequency threshold from the document to be processed, initialize them as sentence nodes, entity mention nodes, and pronoun nodes in a heterogeneous graph, and obtain the corresponding initial feature representations; Using the initialized nodes and their initial feature representations, a heterogeneous graph with enhanced referentiality is constructed, which includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges, and referential-entity soft connection edges. Information from referential nodes is aggregated by soft linking referential nodes and noise is filtered by gating to enhance the feature representation of entity mention nodes. Then, together with the feature representation of sentence nodes, multi-layer message passing is performed in the heterogeneous graph with enhanced referentiality to complete global structural reasoning and update the feature representation of all nodes. Pooling is performed on the feature representations of all sentence nodes to obtain a document-level global vector, which is then used to predict the event type distribution of document samples by a multi-label classifier; Based on the updated feature representations of all nodes, a tracker-based decoder is used to progressively predict event types and corresponding arguments to obtain structured event records.

[0080] It should be noted that each module in the event extraction system of this embodiment corresponds one-to-one with each step in the event extraction method in the above embodiment, and their specific implementation process is the same, so it will not be repeated here.

[0081] The structure of the electronic device according to an embodiment of the present invention will be described in detail below. Figure 9 This is a schematic diagram of the composition structure of an electronic device provided in an embodiment of the present invention. It can be understood that... Figure 9 The diagram shows only an exemplary structure of the electronic device, not the entire structure. Some or all of the structures shown may be implemented as needed.

[0082] The electronic device provided in this embodiment of the invention includes: at least one processor 901, a memory 902, a user interface 903, and at least one network interface 904. The various components in the document-level event extraction model training system or event extraction system are coupled together via a bus system 905. It can be understood that the bus system 905 is used to implement communication between these components. In addition to a data bus, the bus system 905 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 9 The general labeled all buses as Bus System 905.

[0083] The user interface 903 may include a monitor, keyboard, mouse, trackball, click wheel, buttons, touchpad, or touch screen.

[0084] It is understood that memory 902 can be volatile memory or non-volatile memory, or both. In this embodiment of the invention, memory 902 is capable of storing data to support the operation of the terminal. Examples of this data include any computer programs used to operate on the terminal, such as operating systems and applications. The operating system includes various system programs, such as framework layers, core library layers, driver layers, etc., used to implement various basic services and handle hardware-based tasks. Applications can include various applications.

[0085] In some embodiments, the document-level event extraction model training system or event extraction system provided in this invention can be implemented using a combination of hardware and software. For example, the document-level event extraction model training system or event extraction system provided in this invention can be a processor in the form of a hardware decoding processor, programmed to execute the document-level event extraction model training method or event extraction method provided in this invention. For instance, the processor in the form of a hardware decoding processor can employ one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0086] As an example, processor 901 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0087] As an example of the document-level event extraction model training system or event extraction system provided in the embodiments of the present invention, the system provided in the embodiments of the present invention can be directly executed by a processor 901 in the form of a hardware decoding processor. For example, it can be executed by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components to implement the document-level event extraction model training method or event extraction method provided in the embodiments of the present invention.

[0088] The memory 902 in this embodiment of the invention is used to store various types of data to support the operation of the document-level event extraction model training system or the event extraction system, or to store data for execution. Figure 1 or Figure 7The program code for the method shown. Examples of this data include: any executable instructions for operating on a document-level event extraction model training system or an event extraction system, such as executable instructions that can be included in the executable instructions to implement the document-level event extraction model training method or event extraction method of the embodiments of the present invention.

[0089] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including functions for executing... Figure 1 or Figure 7 The program code for the method shown. In such an embodiment, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by the central processing unit, it performs the various functions defined in the apparatus of this application.

[0090] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for training a document-level event extraction model, characterized in that, The method comprises the following steps: extracting sentences, entity mentions and pronouns exceeding a set frequency threshold in a document sample, initializing sentence nodes, entity mention nodes and pronoun nodes corresponding to the sentences, entity mentions and pronouns into a heterogeneous graph and obtaining initial feature representations of the nodes; constructing a pronoun-enhanced heterogeneous graph using the initialized nodes and the initial feature representations of the nodes, the heterogeneous graph including four types of edges, i.e., sentence-sentence edges, sentence-entity edges, entity-entity edges and pronoun-entity soft connection edges; aggregating information of the pronoun nodes through the pronoun-entity soft connection edges, introducing gated filtering noise to enhance feature representations of the entity mention nodes, and performing multi-layer message passing in the pronoun-enhanced heterogeneous graph together with feature representations of the sentence nodes to complete global structure reasoning and update feature representations of all the nodes; pooling feature representations of all the sentence nodes to obtain a document-level global vector, predicting an event type distribution of the document sample through a multi-label classifier, and generating a global event classification task loss; based on the updated feature representations of all the nodes, using a tracker-based decoder to gradually predict event types and corresponding arguments to obtain structured event records and generate an event decoding task loss; summing the event decoding task loss and a weighted global event classification task loss to obtain a total loss of the document-level event extraction model, and jointly optimizing parameters of the document-level event extraction model through multi-task learning. 2.The method of claim 1, wherein, The reference-entity soft connection edge is: through the dynamic calculation of the entity mention node by the gating attention mechanism The association weight between the reference node , forming a virtual semantic edge; the calculation formula is:​ ； ； wherein, is a feature dimension of the key vector; is a query transformation matrix for mapping the feature representation of an entity mention node to a query space; is a feature representation of the th entity mention node; is a transpose operation of a matrix or vector; is a key transformation matrix for mapping the feature representation of a reference node to a key space; is a feature representation of the th reference node; is a normalized exponential function over the reference node dimension ; is a degree of semantic match between an entity mention node and a reference node . 3.The method of claim 1, wherein, In the process of enhancing the feature representations of the entity mention nodes, the entity mention nodes aggregate information of the pronoun nodes through the pronoun-entity soft connection edges, introduce gated filtering noise, and obtain enhanced feature representations of the entity mention nodes: ； ； wherein, is the feature representation of the enhanced entity mention node; denotes a layer normalization operation; is the feature representation of the th entity mention node; is the set of all mention nodes in the referentially augmented heterogeneous graph; is the value transformation matrix for mapping the feature representation of the mention nodes to the semantic value space; is the feature representation of the th mention node; is the association weight between the entity mention node and the mention node ; is the gating coefficient; is the Sigmoid activation function; is the concatenation of the feature representation of the th entity mention node with the feature representation of the th mention node; is the bias feature representation. 4.The method of claim 1, wherein, The initial feature representations of the sentence nodes, the entity mention nodes and the pronoun nodes are obtained using a pre-trained language model. 5.The method of claim 1, wherein, An expression of the total loss of the document-level event extraction model is: ； ； ； wherein, is the total loss; is the event decoding task loss; is the global event classification task loss; is the balancing hyperparameter; is the set of real event records in the document; is the set of event types based on the document predicts the event type and its set of arguments given the document; is the total number of predefined event types; is the label indicating whether the document actually contains the event, 1 for yes, 0 for no; is the probability that the document contains the event.

6. A document-level event extraction model training system, which is obtained by using the steps in the document-level event extraction model training method according to any one of claims 1-5, and comprises: a document preprocessing and node initialization module configured to extract sentences, entity mentions and pronouns exceeding a set frequency threshold in a document sample, initialize sentence nodes, entity mention nodes and pronoun nodes corresponding to the sentences, entity mentions and pronouns into a heterogeneous graph and obtain initial feature representations of the nodes; a pronoun-enhanced heterogeneous graph construction module configured to construct a pronoun-enhanced heterogeneous graph using the initialized nodes and the initial feature representations of the nodes, the heterogeneous graph including four types of edges, i.e., sentence-sentence edges, sentence-entity edges, entity-entity edges and pronoun-entity soft connection edges; a two-stage graph information aggregation module configured to aggregate information of the pronoun nodes through the pronoun-entity soft connection edges, introduce gated filtering noise to enhance feature representations of the entity mention nodes, and perform multi-layer message passing in the pronoun-enhanced heterogeneous graph together with feature representations of the sentence nodes to complete global structure reasoning and update feature representations of all the nodes; a global event classification task loss generation module configured to pool feature representations of all the sentence nodes to obtain a document-level global vector, predict an event type distribution of the document sample through a multi-label classifier, and generate a global event classification task loss; a global event classification task loss generation module configured to pool feature representations of all the sentence nodes to obtain a document-level global vector, predict an event type distribution of the document sample through a multi-label classifier, and generate a global event classification task loss; an event decoding task loss generation module configured to generate an event decoding task loss by using a tracker-based decoder to predict event types and corresponding arguments step by step based on the updated feature representations of all nodes, and obtain structured event records; a total loss construction and parameter optimization module configured to sum the event decoding task loss and the weighted global event classification task loss as a total loss of the document-level event extraction model, and jointly optimize parameters of the document-level event extraction model through multi-task learning.

7. An event extraction method characterized by, The method comprises the following steps: acquiring a document to be processed, which is in the same field as document samples used for training the document-level event extraction model; obtaining structured event records from the document to be processed by using the document-level event extraction model, wherein the document-level event extraction model is trained by using the steps in the document-level event extraction model training method according to any one of claims 1-5; the document-level event extraction model is configured to: extract sentences, entity mentions and pronouns exceeding a set frequency threshold in the document to be processed, and initialize the sentences, entity mentions and pronouns into sentence nodes, entity mention nodes and pronoun nodes of a heterogeneous graph respectively and obtain corresponding initial feature representations; construct a pronoun-enhanced heterogeneous graph by using the initialized nodes and the initial feature representations thereof, wherein the pronoun-enhanced heterogeneous graph comprises four types of edges, i.e., sentence-sentence edges, sentence-entity edges, entity-entity edges and pronoun-entity soft connection edges; aggregate information of the pronoun nodes through the pronoun-entity soft connection edges, introduce a gating filter to filter noise, enhance feature representations of the entity mention nodes, and perform multi-layer message passing on the pronoun-enhanced heterogeneous graph together with the feature representations of the sentence nodes to complete global structure reasoning and update the feature representations of all nodes; perform pooling on the feature representations of all sentence nodes to obtain a document-level global vector, and predict an event type distribution of the document sample by using a multi-label classifier; use a tracker-based decoder to predict event types and corresponding arguments step by step based on the updated feature representations of all nodes, and obtain structured event records.

8. An event extraction system, characterized in that, The method comprises the following steps: a document acquisition module configured to acquire a document to be processed, which is in the same field as document samples used for training the document-level event extraction model; an event extraction module configured to obtain structured event records from the document to be processed by using the document-level event extraction model, wherein the document-level event extraction model is trained by using the steps in the document-level event extraction model training method according to any one of claims 1-5; the document-level event extraction model is configured to: extract sentences, entity mentions and pronouns exceeding a set frequency threshold in the document to be processed, and initialize the sentences, entity mentions and pronouns into sentence nodes, entity mention nodes and pronoun nodes of a heterogeneous graph respectively and obtain corresponding initial feature representations; construct a pronoun-enhanced heterogeneous graph by using the initialized nodes and the initial feature representations thereof, wherein the pronoun-enhanced heterogeneous graph comprises four types of edges, i.e., sentence-sentence edges, sentence-entity edges, entity-entity edges and pronoun-entity soft connection edges; The information of the coreference nodes is aggregated by a coreference- entity soft connection edge, and a gating filter noise is introduced to enhance the feature representation of the entity mention nodes, and then the feature representation of the sentence nodes is combined to perform multi-layer message passing in the coreference-enhanced heterogeneous graph to complete global structure reasoning and update the feature representation of all nodes; The feature representations of all sentence nodes are pooled to obtain a document-level global vector, and a multi-label classifier is used to predict the event type distribution of the document sample; Based on the updated feature representations of all nodes, a tracker-based decoder is used to predict the event type and corresponding arguments step by step to obtain a structured event record.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the document-level event extraction model training method according to any one of claims 1-5; or the program is executed by the processor to implement the steps in the event extraction method according to claim 7.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps in the document-level event extraction model training method according to any one of claims 1-5; or the processor executes the program to implement the steps in the event extraction method according to claim 7.

Images