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

Abstract

The present application relates to the technical field of document-level event extraction, and provides a document-level event extraction model training method, an event extraction method and a system to solve the problems of insufficient remote dependency modeling capability, susceptibility to noise interference and inconsistent event output semantics in the prior art. The document-level event extraction model training method comprises preprocessing a document sample and initializing a node; constructing a reference-enhanced heterogeneous graph; two-stage graph information aggregation; generating a global event classification task loss; generating an event decoding task loss; summing 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 optimizing the parameters of the document-level event extraction model through multi-task learning. It can accurately utilize the pronoun as an explicit semantic bridge to strengthen remote reasoning, effectively filter associated noise, and also integrate global semantic constraints at the document level to ensure consistency of the output.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of document-level event extraction, and particularly relates to a document-level event extraction model training method, an event extraction method and a system. BACKGROUND

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

[0003] Document-level event extraction is one of the core tasks in the field of information extraction, and its goal is to extract complete and structured event records from unstructured whole texts. Unlike the sentence-level task, the documents in real scenarios (such as financial announcements and judicial documents) generally have the phenomenon of "argument scattering": multiple arguments (such as subjects, objects, times and places) of an event are often scattered in different paragraphs or sentences of the document, and a single document often contains multiple associated event instances. This feature requires the model to have strong long-distance semantic association and reasoning ability.

[0004] The traditional breakthrough work Doc2EDAG formulates the event table filling task as a directed acyclic graph sequence generation problem, providing a new idea for end-to-end multi-event extraction. Subsequently, models represented by GIT (Graph-based Interaction Model with a Tracker, Interaction Model with a Tracker) use heterogeneous graph neural networks to aggregate document-level context information by constructing an interaction graph of sentences and entity mentions, further improving performance. However, this kind of method usually models the document as a static set of entity mentions, and its information flow is heavily dependent on the fixed graph structure based on sentence order or adjacency relationship. When the key entity (such as the company name) only appears at the beginning of the document, and the latter uses a large number of pronouns (such as "the company" and "it") for reference, this implicit co-occurrence-based reasoning mode is difficult to establish effective long-distance connections, resulting in low recall rate for scattered arguments.

[0005] Existing document-level relation extraction proves the effectiveness of explicitly modeling anaphora information for improving cross-sentence reasoning by explicitly introducing anaphora nodes and utilizing attention mechanism to assign appropriate weights. However, it is not feasible to directly migrate the anaphora modeling strategy of relation extraction to event extraction. Because event extraction has more intensive argument role assignment and more complex entity interaction, the simple 'full connection first' strategy will introduce a large amount of false anaphora association noise in the graph, causing the node representation to be excessively smoothed. As a result, the model is difficult to distinguish entity clusters belonging to different events, which damages the clarity of event boundaries and the accuracy of argument role discrimination. In addition, the current mainstream event decoder mostly adopts a bottom-up local combination strategy, which lacks explicit awareness of the document macro theme or intent, and is prone to generate pseudo event records that conflict with the overall semantic context. SUMMARY

[0006] To solve the above technical problems, the present application provides a document-level event extraction model training method, an event extraction method and a system, which can accurately use anaphora as an explicit semantic bridge to strengthen long-range reasoning, effectively filter associated noise, and also integrate global semantic constraints at the document level to ensure consistency of the output.

[0007] To achieve the above purpose, the present application adopts the following technical solutions:

[0008] The first aspect of the present application provides a document-level event extraction model training method.

[0009] In one or more embodiments, a document-level event extraction model training method comprises:

[0010] Extracting sentences, entity mentions and anaphora words exceeding a set frequency threshold in a document sample, corresponding to sentence nodes, entity mention nodes and anaphora nodes of a heterogeneous graph and obtaining corresponding initial feature representations;

[0011] Using the initialized nodes and their initial feature representations, a anaphora-enhanced heterogeneous graph is constructed, which includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges and anaphora-entity soft connection edges;

[0012] The information of the anaphora nodes is aggregated through the anaphora-entity soft connection edges and the noise is introduced through the gate filter to enhance the feature representation of the entity mention nodes. Then, the feature representation of the sentence nodes is combined with the feature representation of the sentence nodes in the anaphora-enhanced heterogeneous graph for multi-layer message passing to complete global structure reasoning and update the feature representation of all nodes;

[0013] 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 to generate a global event classification task loss;

[0014] Based on the updated feature representation of all nodes, the tracker-based decoder is used to predict the event type and corresponding argument step by step to obtain the structured event record, and generate the event decoding task loss;

[0015] The event decoding task loss and the weighted global event classification task loss are summed as the total loss of the document-level event extraction model, and the parameters of the document-level event extraction model are jointly optimized through multi-task learning.

[0016] As an implementation, the anaphor-entity soft connection edge is a virtual semantic edge formed by dynamically calculating the association weight between the entity mention node and the anaphor node through a gated attention mechanism, and its calculation formula is:

[0017] ;

[0018] ;

[0019] wherein, is the feature dimension of the key vector; is a query transformation matrix used to map the feature representation of the entity mention node to the query space; is the feature representation of the th entity mention node; is the transpose operation of the matrix or vector; is a key transformation matrix used to map the feature representation of the anaphor node to the key space; is the feature representation of the th anaphor node; is a normalized exponential function in the dimension of the anaphor node ; is the degree of semantic matching between the entity mention node and the anaphor node .

[0020] As an implementation, in the process of enhancing the feature representation of the entity mention node, the entity mention node aggregates the information of the anaphor node through the anaphor-entity soft connection edge, introduces gated filtering noise, and obtains the enhanced feature representation of the entity mention node:

[0021] ;

[0022] ;

[0023] wherein, is the enhanced feature representation of the entity mention node; ​representation layer normalization operation; feature representation of the th entity mention node; a set of all anaphora nodes in the augmented heterogeneous graph; a value transformation matrix for mapping the feature representation of the anaphora node to a semantic value space; feature representation of the th anaphora node; an association weight between the entity mention node and the anaphora node ; a gating coefficient; a Sigmoid activation function; concatenating the feature representation of the th entity mention node with the feature representation of the th anaphora node; a bias feature representation.

[0024] As an implementation, the initial feature representations of the sentence node, the entity mention node and the anaphora node are obtained by using a pre-trained language model.

[0025] As an implementation, the expression of the total loss of the document-level event extraction model is as follows:

[0026] ;

[0027] ;

[0028] ;

[0029] wherein, the total loss; the event decoding task loss; the global event classification task loss; a balance hyperparameter; a set of real event records in the document; a conditional probability of predicting the event type and the argument set thereof based on the document ; a total number of predefined event types; a label indicating whether the document actually contains the th event, 1 for containing and 0 for not containing; a probability of predicting that the document contains the th event.

[0030] The second aspect of the present application provides a document-level event extraction model training system.​

[0031] In one or more embodiments, a document-level event extraction model training system comprises:

[0032] a document preprocessing and node initialization module for extracting sentences, entity mentions and pronouns exceeding a set frequency threshold in a document sample, initializing the sentences, entity mentions and pronouns into a heterogeneous graph and obtaining initial feature representations of the nodes;

[0033] a pronoun-enhanced heterogeneous graph construction module for constructing a pronoun-enhanced heterogeneous graph containing four types of edges, i.e., sentence-sentence edges, sentence-entity edges, entity-entity edges and pronoun-entity soft connection edges, using the initialized nodes and their initial feature representations;

[0034] a two-stage graph information aggregation module for 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 the feature representations of the sentence nodes to complete global structure reasoning and update feature representations of all nodes;

[0035] a global event classification task loss generation module for obtaining a document-level global vector by pooling feature representations of all sentence nodes, predicting an event type distribution of the document sample through a multi-label classifier and generating a global event classification task loss;

[0036] an event decoding task loss generation module for generating an event decoding task loss by using a tracker-based decoder to step-by-step predict event types and corresponding arguments based on the updated feature representations of all nodes and obtaining structured event records;

[0037] a total loss construction and parameter optimization module for summing the event decoding task loss and a weighted global event classification task loss as 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.

[0038] A third aspect of the present application provides an event extraction method.

[0039] In one or more embodiments, an event extraction method comprises:

[0040] obtaining a document to be processed, which is in the same field as the document sample used to train the document-level event extraction model;

[0041] 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 described above;

[0042] The document-level event extraction model is configured to:

[0043] extracting sentences, entity mentions and pronouns exceeding a set frequency threshold in the document to be processed, initializing the sentences, entity mentions and pronouns into a heterogeneous graph as sentence nodes, entity mention nodes and pronoun nodes and obtaining corresponding initial feature representations;

[0044] using the initialized nodes and the initial feature representations thereof, constructing a pronoun-enhanced heterogeneous graph containing four types of edges, namely, sentence-sentence edges, sentence-entity edges, entity-entity edges and pronoun-entity soft connection edges;

[0045] aggregating information of the pronoun nodes through the pronoun-entity soft connection edges and introducing gated filtering noise to enhance feature representations of the entity mention nodes, and then 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 nodes;

[0046] pooling feature representations of all sentence nodes to obtain a document-level global vector, and predicting an event type distribution of the document sample through a multi-label classifier;

[0047] based on the updated feature representations of all nodes, using a tracker-based decoder to gradually predict event types and corresponding arguments to obtain a structured event record.

[0048] A fourth aspect of the present application provides an event extraction system.

[0049] In one or more embodiments, an event extraction system comprises:

[0050] a document acquisition module configured to acquire a document to be processed, which is in the same field as document samples used for training a document-level event extraction model;

[0051] an event extraction module configured to use the document-level event extraction model to obtain a structured event record from the document to be processed; wherein the document-level event extraction model is trained using the steps of the document-level event extraction model training method described above;

[0052] The document-level event extraction model is configured to:

[0053] extracting sentences, entity mentions and pronouns exceeding a set frequency threshold in the document to be processed, initializing the sentences, entity mentions and pronouns into a heterogeneous graph as sentence nodes, entity mention nodes and pronoun nodes and obtaining corresponding initial feature representations;

[0054] using the initialized nodes and the initial feature representations thereof, constructing a pronoun-enhanced heterogeneous graph containing four types of edges, namely, sentence-sentence edges, sentence-entity edges, entity-entity edges and pronoun-entity soft connection edges;

[0055] The information of the reference-entity soft connection edge is aggregated to the reference node, and a gating filter is introduced to filter noise, so as to enhance the feature representation of the entity mention node, and then the feature representation of the sentence node is combined in the reference-enhanced heterogeneous graph to perform multi-layer message passing, complete global structure reasoning, and update the feature representation of all nodes;

[0056] 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;

[0057] Based on the updated feature representation of all nodes, a tracker-based decoder is used to predict the event type and corresponding argument step by step to obtain a structured event record.

[0058] The fifth aspect of the present application provides a computer-readable storage medium.

[0059] In one or more embodiments, a computer-readable storage medium has a computer program stored thereon, and the program is executed by a processor to implement the steps of the document-level event extraction model training method described above.

[0060] In other embodiments, a computer-readable storage medium has a computer program stored thereon, and the program is executed by a processor to implement the steps of the event extraction method described above.

[0061] The sixth aspect of the present application provides an electronic device.

[0062] In one or more embodiments, an electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the program to implement the steps of the document-level event extraction model training method described above.

[0063] In other embodiments, an electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the program to implement the steps of the event extraction method described above.

[0064] Compared with the prior art, the present application has the following advantages:

[0065] (1) The application constructs a reference-enhanced heterogeneous graph in the process of document-level event extraction model training, aggregates the information of reference-entity soft connection edges according to the reference-entity soft connection edges, introduces a gating filter noise, enhances the feature representation of the entity mention node, can explicitly filter and establish a high-confidence reference correlation, and transmits the enhanced feature representation of the entity mention node and the feature representation of the sentence node through multi-layer message passing, completes global structure reasoning, updates the feature representation of all nodes, thereby constructing a cross-sentence semantic bridge and suppressing the propagation of irrelevant noise; 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, the global event constraint is comprehensively realized, and the semantic supervision of the document level is injected by means of the multi-task learning mechanism, and the internal consistency of the event record is improved.

[0066] (2) The application uses the trained document-level event extraction model to perform document-level event extraction on the documents in the same field as the document samples used to train the document-level event extraction model, and in the process of document-level event extraction, a reference-enhanced heterogeneous graph is constructed by introducing a reference node and a gating attention, effectively solving the problem of long-range reasoning break caused by implicit reference chain in long documents, which can accurately use the reference word as an explicit semantic bridge to strengthen long-range reasoning, and effectively filter associated noise. At the same time, the internal consistency of the event record is realized. BRIEF DESCRIPTION OF DRAWINGS

[0067] The drawings accompanying the specification of the application form part of the application and serve to provide further understanding of the application, the illustrative embodiments of the application and their description serve to explain the application without constituting an improper limitation thereof.

[0068] Figure 1 is a document-level event extraction model training method flowchart of an embodiment of the application;

[0069] Figure 2 is a document-level event extraction model overall architecture of an embodiment of the application;

[0070] Figure 3 is a systematic experimental comparison of the document-level event extraction model and the traditional GIT model on the public benchmark dataset ChFinAnn of an embodiment of the application;

[0071] Figure 4 is a comparison of the document-level event extraction model and the baseline model under different span parameters of an embodiment of the application;

[0072] Figure 5 is a case of the gating attention mechanism of an embodiment of the application;

[0073] Figure 6 is a document-level event extraction model training system structure schematic diagram of an embodiment of the application;

[0074] Figure 7 is a flow chart of an event extraction method according to an embodiment of the present application;

[0075] Figure 8 is a schematic diagram of an event extraction system according to an embodiment of the present application;

[0076] Figure 9 is a schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0077] The present application will be further described with reference to the drawings and embodiments.

[0078] It should be noted that the following detailed description is merely exemplary and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0079] It should be noted that the terms used herein are merely for the purpose of describing specific embodiments and are not intended to limit exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should be further understood that the terms "comprise" and / or "include" as used herein indicate the presence of a feature, step, operation, device, component and / or a combination thereof.

[0080] Graph neural networks are widely used in document-level information extraction due to their strong structure modeling capability. In the field of relation extraction, the LSR model (Latent Space Representation Model) uses document structure as a latent variable for reasoning, and the GAIN model (which realizes the fusion of heterogeneous information in the document by constructing a double-layer graph structure of mention level graph (hMG) and entity level graph (EG)) constructs a double-layer graph of mentions and entities to fuse heterogeneous information. Inspired by this, the GIT model first introduced heterogeneous graph interaction into document-level event extraction, and constructed a fully connected graph between 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, and it fails to fully utilize explicit linguistic cues such as pronouns to construct more accurate semantic paths.

[0081] Anaphora resolution is the core task of connecting scattered references in a text. The existing technology innovatively takes the pronoun as a node in the graph and enhances the relationship reasoning by connecting all potential anaphoric pairs, which proves the effectiveness of explicit anaphora modeling. However, event extraction is more complex than relation extraction, and it is necessary to identify multiple specific arguments from a large number of candidate entities. In the event document with dense arguments, the dense connection strategy will introduce significant noise, leading to over-smoothing of the graph node representation. Embodiments of the present invention introduce a learnable gating mechanism to dynamically filter low-quality anaphoric associations and retain high-confidence semantic bridges.

[0082] Document-level event extraction aims to extract structured event records from complete documents, and the difficulty lies in "argument dispersion", that is, the arguments of a single event often span multiple distant sentences, which has high requirements for the cross-sentence reasoning ability of the model. The existing heterogeneous graph-based model mostly relies on the physical proximity between sentences for implicit information transmission, and cannot effectively utilize the explicit semantic association represented by the pronoun, resulting in insufficient long-range dependency modeling ability and being easily disturbed by noise; at the same time, the mainstream bottom-up decoding mechanism lacks explicit constraints on the overall theme of the document, and is prone to produce semantically inconsistent event outputs. To solve the above problems, the embodiments of the present invention construct a reference-enhanced heterogeneous graph structure (i.e. GA-GIT, i.e. Graph Attention Graph-based Interaction Model with a Tracker architecture), which can explicitly filter and establish high-confidence anaphoric associations through the designed gating attention mechanism, thereby constructing cross-sentence semantic bridges and suppressing the propagation of irrelevant noise. In addition, a global event constraint module is introduced, which injects semantic supervision at the document level through a multi-task learning mechanism to improve the internal consistency of event records. Experiments on the ChFinAnn dataset in the financial field show that the GA-GIT architecture of the embodiments of the present invention achieves a Micro-F1 value of 82.2%, which is significantly better than the existing mainstream baseline model, and the recall rate is particularly improved in the long-distance extraction scenario with highly dispersed arguments.

[0083] Figure 1 The principle diagram of the document-level event extraction model training method of the embodiments of the present invention is given. Combined with Figure 1 and Figure 2 The document-level event extraction model training method of the embodiments of the present invention can include the following steps S101-S106.

[0084] The specific implementation process of steps S101-S106 is as follows:

[0085] Step S101: Extract sentences, entity mentions, and pronouns in the document sample that exceed the set frequency threshold, initialize them as sentence nodes, entity mention nodes, and reference nodes of the heterogeneous graph respectively, and obtain the corresponding initial feature representations.

[0086] Given a document containing sentences, the goal of document-level event extraction is to extract a set of event records . Each event record consists of a predefined event type and its corresponding set of arguments , where represents the argument role type, is an entity mention in the document. In addition, to explicitly model cross-sentence reference, the embodiments of the present invention define a set of reference candidates [[ID=2l]]extracted from the document as a potential semantic hub for connecting scattered arguments; is the total number of pronouns extracted; is the set of event types.

[0087] In the specific implementation process, the document sample is segmented into sentences and entities are recognized, and pronouns in the document that exceed the set frequency threshold are extracted; each sentence, entity mention, and pronoun is initialized as a sentence node, entity mention node, and reference node in the heterogeneous graph respectively, and the initial feature representations of each node are obtained using a pre-trained language model.

[0088] 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 in the document that exceed the set frequency threshold.

[0089] 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;

[0090] 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 reference nodes.

[0091] 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.

[0092] Expand the document into a Token sequence and input into a field-adapted pre-trained language model (e.g., Mengzi-BERT-base-fi model) to obtain context embeddings The matrix contains the feature representation of all Tokens in the document . Among them, represents the length of the Token sequence input into the pre-trained language model, that is, the total number of Tokens contained after the document is processed by 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.

[0093] The initial feature representations of the sentence nodes and entity mention nodes are obtained through max-pooling operations:

[0094] ;

[0095] ;

[0096] Among them, is the full-document context embedding matrix output by the pre-trained language model with a dimension of ; and are the starting and ending index positions of the th sentence node in the document sequence; and are the starting and ending index positions of the th entity mention node in the document sequence; is the max-pooling operation along the sequence length dimension, which is used to aggregate the variable-length Token vector sequence into a fixed-dimension feature vector; and are the initial feature vectors of the th sentence node and the th entity mention node after aggregation, both with a dimension of .

[0097] According to the initial feature representations of the sentence nodes and entity mention nodes , the sentence node feature set and the entity node feature set are obtained.

[0098] 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.

[0099] 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:

[0100] (1) Sentence-Sentence Edge : Based on sentence adjacency or discourse structure connections;

[0101] (2) Sentence-entity edge : Connects a sentence to its contained entity references;

[0102] (3) Entity-Entity Edge : Connects entities within the same sentence or those with explicit coreference;

[0103] (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:

[0104] ;

[0105] ;

[0106] 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 semantic matching degree between them. The greater the value is, the higher the possibility that the referential node refers to the entity mention node is.

[0107] Step S103: The information of the referential node is aggregated through the referential-entity soft connection edge, and the gating filtering noise is introduced to enhance the feature representation of the entity mention node. Then, the multi-layer message passing is performed in the referential-enhanced heterogeneous graph with the feature representation of the sentence node to complete the global structure reasoning and update the feature representation of all nodes.

[0108] In the process of enhancing the feature representation of the entity mention node, the entity mention node aggregates the information of the referential node through the referential-entity soft connection edge, and introduces the gating filtering noise to obtain the enhanced feature representation of the entity mention node:

[0109] ;

[0110] ;

[0111] wherein, is the enhanced feature representation of the entity mention node; represents a layer normalization operation; is the feature representation of the th entity mention node; is the set of all referential nodes in the referential-enhanced heterogeneous graph; is a value transformation matrix for mapping the feature representation of the referential node to a semantic value space; is the feature representation of the th referential node; is the association weight between the entity mention node and the referential node ; is a gating coefficient calculated by a Sigmoid activation function, and the value range is between 0 and 1, which is used to explicitly filter noise; when tends to 0, it means that the model judges that the connection is noise and blocks it. is a Sigmoid activation function; is the feature representation of the th entity mention node spliced with the feature representation of the th referential node; is a bias feature representation.

[0112] The entity mention node aggregates the information of the referential node through the gating soft connection edge, realizes the cross-sentence semantic enhancement, and splices the enhanced entity node with the sentence node on the sentence-sentence edge​ , sentence-entity edges , and entity-entity edges Multi-round message passing is performed on the graph to complete global structure inference and update features of all nodes.

[0113] Step S105: Pooling the feature representations of all sentence nodes to obtain a document-level global vector, and obtaining a structured event record through multi, to generate an event decoding task loss.

[0114] To improve the global consistency of event type prediction, the embodiment of the application adds a document-level event type classification auxiliary task. The document global vector is obtained by maximum pooling all sentence representations , and then the event type distribution of the document is predicted by a classifier: ;

[0115] The binary cross-entropy loss of this task serves as a global constraint signal.

[0116] wherein, represents the predicted document-level event type probability distribution vector; represents a Sigmoid activation function; represents a weight matrix of the classifier; represents a bias vector of the classifier.

[0117] Step S106: Summing the event decoding task loss and the weighted global event classification task loss to obtain the total loss of the document-level event extraction model, and then jointly optimizing the parameters of the document-level event extraction model through multi-task learning.

[0118] The decoder adopts a sequence generation strategy based on Tracker to generate event types and arguments step by step, and its loss is denoted as .

[0119] The expression of the total loss of the document-level event extraction model is:

[0120] ;

[0121] ;

[0122] ;

[0123] wherein, is the total loss; is the event decoding task loss; is the global event classification task loss; is a balance hyperparameter used to adjust the contribution degree 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.

[0124] Data sets and metrics:

[0125] 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 reason across sentences in long documents.

[0126] 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.

[0127] 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.

[0128] This invention compares GA-GIT with the following mainstream document-level event extraction models:

[0129] DCFEE: An early pipeline model that completes arguments through center sentence detection and neighbor expansion strategies.

[0130] Doc2EDAG: A Transformer-based end-to-end sequence generation model that models the extraction task as EDAG path generation.

[0131] GIT: The current SOTA model that utilizes heterogeneous graph networks and a Tracker mechanism for extraction. GA-GIT follows its base architecture but improves on the graph interaction mechanism.

[0132] Table 1 shows the comparative results of different models on the test set.

[0133] Table 1 shows the comparative results of different models on the test set.

[0134]

[0135] From Table 1, it can be observed that in terms of comprehensive performance improvement, the GA-GIT of the embodiment of the present application achieves the latest optimal performance, with a Micro-F1 value of 82.2%, which is 1.9 percentage points higher than the strong baseline GIT, verifying the effectiveness of the anaphora enhancement strategy of the GA-GIT of the embodiment of the present application.

[0136] Significant improvement in recall: It is worth noting that the improvement in recall (+3.4%) of GA-GIT is more significant than the improvement in precision (+0.5%), which supports the core hypothesis of the embodiment of the present application: existing graph-based methods (such as GIT) often miss arguments due to the breakage of implicit connections. In contrast, the anaphora-enhanced heterogeneous graph of the embodiment of the present application establishes an explicit semantic bridge through anaphors, successfully recovering these "hard-to-find" arguments (especially subject entities through pronoun anaphora), thereby significantly improving the recall rate without sacrificing precision.

[0137] To explore the contribution of each component in the GA-GIT of the embodiment of the present application, the embodiment of the present application conducts an ablation experiment by removing specific modules, and the results are reported in Table 2.

[0138] Table 2 shows the results of the ablation experiment.

[0139]

[0140] The experimental analysis is as follows:

[0141] Necessity of anaphora nodes: Removing anaphora nodes (w / o Anaphor Nodes) results in the largest performance decline (-1.7%), proving that explicitly modeling anaphoric information is crucial for capturing cross-sentence dependencies, and relying solely on physical adjacency edges cannot effectively convey long-distance semantics.

[0142] Denoising effect of gating mechanism: If the gating mechanism is removed (w / o Gating), the performance drops by 0.9%. This confirms that in argument-dense event graphs, simple full connection will introduce a lot of false coreference noise, and the gating mechanism effectively acts as a "semantic filter" to prevent over-smoothing of the graph.

[0143] Correcting effect of global constraint: Removing the global constraint (w / o Global Constraint) leads to a 0.6% drop in F1. Further analysis finds that this variant is prone to predicting mutually exclusive event types in multi-event documents, indicating that global semantic supervision helps to regulate the behavior of the decoder and ensure that local predictions are consistent with the main theme of the document.

[0144] To further verify the model's ability to handle the "argument dispersion" problem, the embodiments of the present application follow the GIT setting and divide the test set samples into four difficulty-increasing groups (sets I, II, III, IV) according to the number of sentences spanned by event arguments. Among them, set I (1 sentence) represents the simplest scenario with the most concentrated arguments, and set IV represents the most difficult scenario with the largest argument span (more than 10 sentences).

[0145] As shown in Figure 4 , the embodiments of the present application compare the F1 performance of GA-GIT and GIT in different groups: it can be seen that the baseline model (red dotted line) significantly deteriorates in long-distance (set III and set IV) scenarios. In contrast, GA-GIT (blue solid line) shows greater stability. In particular, in the set IV group with a distance greater than 10, GA-GIT achieves a significant improvement of 4.0%, proving its effectiveness in capturing long-distance context information.

[0146] In the simple set I, the performance gap between the two is small. As the degree of argument dispersion increases (sets II and IV), the performance of GIT deteriorates significantly, indicating that implicit graph connections gradually fail at long distances. GA-GIT exhibits strong robustness in the most challenging set IV, achieving a 4.0% F1 improvement over GIT. This trend intuitively proves that the reference node acts as a "semantic relay station", effectively shortening the reasoning path and solving the problem of long-distance dependency breakage.

[0147] To visually demonstrate the working mechanism of GA-GIT, the embodiments of the present application select a specific test sample for visual analysis, as shown in Figure 5 . The gating mechanism assigns a high weight (g = 0.97) to the correct antecedent, thereby establishing a strong semantic bridge, while suppressing noise links with irrelevant entities with a low weight (g = 0.03).

[0148] In this case, the core entity "X Co. Ltd." only appears in the first sentence of the document, and the subsequent 6th sentence refers to the subject through the pronoun "it" and describes the "reduction" action. The baseline model GIT fails to establish the connection between the 1st sentence and the 6th sentence, mistakenly omitting the argument of the shareholder. In contrast, the GA-GIT successfully identifies the pronoun "it" as a key node. Through the gating attention mechanism, the model establishes a high-weight virtual edge between "it" and "X Co. Ltd.", cutting through the shortcut of semantic transmission, and correctly extracting the complete event record. Figure 3 (a) the traditional GIT model in the prior art in the systematic experiment on the public benchmark data set ChFinAnn; Figure 3 (b) in the prior art in the systematic experiment on the public benchmark data set ChFinAnn; by comparison, the document-level event extraction model of the embodiment of the present application can accurately use pronouns as explicit semantic bridges, and strengthen long-range reasoning.

[0149] As shown in Figure 6 The document-level event extraction model training system provided by the embodiment of the present application can be realized in a software manner. The document-level event extraction model training system includes the following software modules:

[0150] The document preprocessing and node initialization module 601 is used to extract sentences, entity mentions and pronouns exceeding a set frequency threshold in the document sample, and corresponds to the sentence nodes, entity mention nodes and pronoun nodes of the heterogeneous graph initialized and obtains the corresponding initial feature representation;

[0151] The pronoun-enhanced heterogeneous graph construction module 602 is used to construct a pronoun-enhanced heterogeneous graph using the initialized nodes and their initial feature representations, which includes four types of edges: sentence-sentence edges, sentence-entity edges, entity-entity edges and pronoun-entity soft connection edges;

[0152] The two-stage graph information aggregation module 603 is used to aggregate the information of the pronoun nodes through the pronoun-entity soft connection edges and introduce a gating filter to filter noise, to enhance the feature representation of the entity mention nodes, and then perform multi-layer message passing with the feature representation of the sentence nodes in the pronoun-enhanced heterogeneous graph to complete global structure reasoning and update the feature representation of all nodes;

[0153] The global event classification task loss generation module 604 is used to pool the feature representation of all sentence nodes to obtain a document-level global vector, predict the event type distribution of the document sample through a multi-label classifier, and generate a global event classification task loss;

[0154] an event decoding task loss generation module 605 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;

[0155] a total loss construction and parameter optimization module 606 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.

[0156] It should be noted that each module in the document-level event extraction model training system of the embodiment of the present application corresponds to each step in the document-level event extraction model training method of the above-described embodiment one by one, and the specific implementation process is the same, which will not be described here again.

[0157] In one or more embodiments, as shown in Figure 7 an event extraction method, comprising:

[0158] Step S701: obtaining a document to be processed, which is in the same field as the document sample used to train the document-level event extraction model;

[0159] Step S702: 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 described above;

[0160] The document-level event extraction model is configured to:

[0161] extracting sentences, entity mentions and anaphors exceeding a set frequency threshold in the document to be processed, and initializing the sentences, entity mentions and anaphors into sentence nodes, entity mention nodes and anaphor nodes of a heterogeneous graph and obtaining corresponding initial feature representations;

[0162] constructing an anaphor-enhanced heterogeneous graph by using the initialized nodes and their initial feature representations, the anaphor-enhanced heterogeneous graph including four types of edges, i.e., sentence-sentence edges, sentence-entity edges, entity-entity edges and anaphor-entity soft connection edges;

[0163] aggregating information of the anaphor nodes through the anaphor-entity soft connection edges and introducing a gating filter to filter noise, so as to enhance the feature representations of the entity mention nodes, and then performing multi-layer message passing on the anaphor-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;

[0164] pooling the feature representations of all sentence nodes to obtain a document-level global vector, and predicting an event type distribution of the document sample by using a multi-label classifier;

[0165] Based on the updated feature representation of all nodes, a tracker-based decoder is used to predict the event type and corresponding arguments step by step to obtain the structured event record.

[0166] In one or more embodiments, as shown in Figure 8 An event extraction system includes:

[0167] A document acquisition module 801 is configured to acquire a document to be processed, which is in the same field as the document sample used to train the document-level event extraction model;

[0168] An event extraction module 802 is configured to obtain a structured event record 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 described above.

[0169] The document-level event extraction model is configured to:

[0170] 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 and obtain corresponding initial feature representations;

[0171] Using the initialized nodes and their initial feature representations, a pronoun-enhanced heterogeneous graph is constructed, which includes four types of edges, namely sentence-sentence edges, sentence-entity edges, entity-entity edges and pronoun-entity soft connection edges.

[0172] The information of the pronoun nodes is aggregated through the pronoun-entity soft connection edges, and a gating filter is introduced to filter noise, so as 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 pronoun-enhanced heterogeneous graph, complete global structure reasoning, and update the feature representation of all nodes.

[0173] The feature representations of all sentence nodes are pooled to obtain a document-level global vector, and the event type distribution of the document sample is predicted by using a multi-label classifier.

[0174] Based on the updated feature representation of all nodes, a tracker-based decoder is used to predict the event type and corresponding arguments step by step to obtain the structured event record.

[0175] It should be noted that each module in the event extraction system of the embodiment of the present application corresponds to each step in the event extraction method of the above-mentioned embodiment, and the specific implementation process is the same, which will not be repeated here.

[0176] The structure of the electronic device of the embodiment of the present application will be described in detail below, Figure 9 The composition structure diagram of the electronic device provided by the embodiment of the present application can be understood,Figure 9 Only exemplary structures of the electronic device are shown, not all structures, and part or all of the shown structures can be implemented as needed.

[0177] The electronic device provided by the embodiment of the present application 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 the event extraction system are coupled together through a bus system 905. It can be understood that the bus system 905 is used to realize the connection communication between the components. In addition to including a data bus, the bus system 905 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, all the buses are marked as the bus system 905 in Figure 9 .

[0178] The user interface 903 can include a display, a keyboard, a mouse, a trackball, a click wheel, a key, a button, a touchpad, or a touch screen, etc.

[0179] It can be understood that the memory 902 can be a volatile memory or a non-volatile memory, and can also include both volatile and non-volatile memories. The memory 902 in the embodiment of the present application can store data to support the operation of the terminal. Examples of these data include any computer programs for operating on the terminal, such as an operating system and an application program. The operating system contains various system programs, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks. The application program can include various application programs.

[0180] In some embodiments, the document-level event extraction model training system or the event extraction system provided by the embodiment of the present application can be implemented in a combination of software and hardware. As an example, the document-level event extraction model training system or the event extraction system provided by the embodiment of the present application can be a processor in the form of a hardware decoding processor, which is programmed to execute the document-level event extraction model training method or the event extraction method provided by the embodiment of the present application. For example, the processor in the form of a hardware decoding processor can use 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 elements.

[0181] As an example, the processor 901 can be an integrated circuit chip having a processing capability of signals, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor.

[0182] As an example, the processor 901 can be an integrated circuit chip having a processing capability of signals, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc., wherein the general-purpose processor can be a microprocessor or any conventional processor.

[0183] The memory 902 in the embodiment of the present application 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 store program codes for executing the method shown in the method. Figure 1 or Figure 7 The examples of these data include any executable instructions for operating on the document-level event extraction model training system or the event extraction system, such as executable instructions, and the program implementing the document-level event extraction model training method or the event extraction method of the embodiment of the present application can be included in the executable instructions.

[0184] In particular, according to the embodiment of the present application, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the embodiment of the present application includes a computer program product including a computer program carried on a computer readable medium, and the computer program includes program codes for executing the method shown in the method. Figure 1 or Figure 7 In such an embodiment, the computer program can be downloaded and installed from a network by a communication part, and / or installed from a detachable medium. When the computer program is executed by the central processing unit, various functions defined in the device of the present application are executed.

[0185] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart and / or block diagram in the variation of the present application can be implemented by computer program instructions. These program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing apparatus, or other programmable data processing apparatus to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing apparatus, generate a means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 The flowchart and / or block diagram in the variation of the present application can be implemented by computer program instructions. These program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing apparatus, or other programmable data processing apparatus to produce a machine, so that the instructions, which are executed via the processor of the computer or other programmable data processing apparatus, generate a means for implementing the functions specified in the flowchart and / or block diagram block or blocks.

[0186] The above merely provides the preferred embodiment of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the protection scope of the present application.

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 a feature representation of the enhanced entity mention node; denotes a layer normalization operation; is a feature representation of the th entity mention node; is a set of all mention nodes in the coreference-enhanced heterogeneous graph; is a value transformation matrix for mapping the feature representation of a mention node to a semantic value space; is a feature representation of the th mention node; is an association weight between the entity mention node and the mention node ; is a gating coefficient; is a Sigmoid activation function; is a concatenation of the feature representation of the th entity mention node and the feature representation of the th mention node; is a 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 of 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