Document-level event argument extraction method based on role focus heterogeneous graph

Abstract

The invention provides a document-level event argument extraction method based on a role focus heterogeneous graph. The method comprises the steps that event types and role nodes are associated; associating a trigger word and a central sentence in the event instance, and constructing a role focus heterogeneous graph; inputting the role focus heterogeneous graph into a graph attention network for learning, and obtaining node representation after the graph attention network learning; performing document word capture on the node representation after graph attention network learning by utilizing a pre-training language model, and performing weighted fusion on the node representation after graph attention network learning; the invention develops a document-level event argument extraction model based on a role focus heterogeneous graph. By using a graph neural network, role representation updated through a role focus heterogeneous graph is obtained, and an argument span serving as a role is obtained by adopting a span selection strategy based on the role.

Description

Technical Field

[0001] The present invention relates to the technical field of information extraction, and in particular to a document-level event argument extraction method based on a role-focus heterogeneous graph. Background Art

[0002] Since existing document-level event recognition and argument extraction research all deal with entity arguments, in order to expand the scope of application of event extraction, a large number of scholars have focused on document-level event argument extraction research of complex arguments. Complex arguments refer to complex objects described by a text that act as event arguments. They may be text fragments composed of words (phrases) or entities with restrictive descriptive content, or text fragments composed of multiple entities in parallel, or a text fragment that may contain event information.

[0003] Most of the traditional classification strategies in existing research use exhaustive methods to determine candidate arguments, which results in a huge number of candidate arguments, thus affecting the argument extraction effect; machine reading comprehension strategies and text generation strategies do not explicitly capture the associated semantics between roles; and the existing span selection strategy models do not fully explore and utilize the roles and their associated semantics contained in event patterns and event instances. Summary of the invention

[0004] In view of the above situation, the main purpose of the present invention is to propose a document-level event argument extraction method based on a role-focus heterogeneous graph to solve the above technical problems.

[0005] The present invention proposes a document-level event argument extraction method based on a role-focus heterogeneous graph, the method comprising the following steps: Step 1: associate the event type and the role node to obtain the association semantics of the associated role in the event pattern; Step 2: Associating the trigger words and the central sentence in the event instance, obtaining the associated semantics of the trigger words and the central sentence in the event instance, associating the associated semantics of the associated roles in the event pattern and the associated semantics of the trigger words and the central sentence in the event instance, and constructing a role focus heterogeneous graph; Step 3: Input the role focus heterogeneous graph into the graph attention network for learning, and obtain the node representation after learning by the graph attention network; Step 4: Use the pre-trained language model to capture document terms from the node representation learned by the graph attention network, and perform weighted fusion with the node representation learned by the graph attention network to obtain the fused semantic representation; Step 5: Construct an event argument extraction model based on the role focus heterogeneous graph, and use the event argument extraction model combined with the fused semantic representation to perform cross-role argument extraction to obtain the probability of the role's span start position on the document and the probability of the role's span end position on the document; The loss function of the model is constructed based on the probability of the character's span starting position on the document and the probability of the character's span ending position on the document; Optimize the event argument extraction model based on the loss function of the model to obtain the optimized event argument extraction model; Step 6: Use the optimized event argument extraction model to obtain the event argument results.

[0006] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention designs a role semantics capture strategy for a role focus heterogeneous graph that reflects the semantics of event structure. It captures the semantics associated with roles, event types, trigger words, and central sentences from two levels: event patterns and event instances, enriches role semantics around event structures, and improves the effect of span selection based on roles; 2. This paper develops a document-level event argument extraction model based on a role-focus heterogeneous graph. By using a graph neural network, the role representation updated by the role-focus heterogeneous graph is obtained, and the span of the argument that plays the role is obtained by adopting a role-based span selection strategy.

[0007] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a flowchart of the steps of the document-level event argument extraction method based on the role-focus heterogeneous graph proposed by the present invention.

[0009] Figure 2 Schematic diagram of the model framework of the document-level event argument extraction method based on role-focus heterogeneous graph proposed in the present invention. DETAILED DESCRIPTION

[0010] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout are the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0011] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to provide some ways to implement the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0012] See also Figure 1The embodiment of the present invention proposes a document-level event argument extraction method based on a role-focus heterogeneous graph, the method comprising the following steps: Step 1: Associate the event type and the role node to obtain the associated semantics of the associated role in the event pattern.

[0013] In step 1, the event type and the role node are associated to obtain the associated semantics of the associated role in the event pattern, which specifically includes the following steps: Based on the event pattern, the event type and role nodes are represented by sets, and the event type set and role set are obtained respectively; Establish an edge between the first event type and the second event type to obtain the association semantics between the first event type and the second event type; Establish an edge between the first event type and the first role node to obtain the association semantics between the first event type and the first role node; An edge is established between the first role node and the second role node to obtain the association semantics between the first role node and the second role node.

[0014] In the steps of respectively expressing the event type and the role node as a set and obtaining the set of event types and the set of roles, the corresponding processes have the following relational expressions: ; in, Represents a set of event types. Indicates the first event type, The node number indicating the event type. Indicates the number of event types contained in the document. Indicates Event Type A collection of characters included, Represents the first role node, Indicates The event type contains role nodes, Indicates the number of roles included in the event type; In the step of establishing an edge between the first event type and the second event type to obtain the association semantics between the first event type and the second event type, the relationship between the corresponding process is as follows: ; in, represents the adjacency matrix corresponding to the role-focus heterogeneous graph, represents the row or column of the corresponding node in the adjacency matrix, Indicates Event Type The character collection includes: Indicates Event Type Included character collection; In the step of establishing an edge between the first event type and the first role node to obtain the association semantics between the first event type and the first role node, the relationship between the corresponding process is as follows: ; in, Indicates role nodes; In the step of establishing an edge between the first role node and the second role node to obtain the association semantics between the first role node and the second role node, the relationship between the corresponding process is as follows: ; in, Indicates A role node.

[0015] Furthermore, an event type-event type edge is established to connect event type nodes containing the same role name. Different event types may share roles with the same name, and this type edge establishes an association between different event types. Establish an event type-role edge, connecting the event type node with each role node it contains. Different event types have different event templates, involving different sets of roles. Through this type of edge, the semantic information of the event type can be transferred to the corresponding role.

[0016] Establish role-role edges, and connect all role nodes under the same event type. Different roles under the same event type are often used to describe different aspects of the same type of event. This type of edge can capture the semantics of the association between event roles and roles.

[0017] Step 2: Associate the trigger words and the central sentence in the event instance, obtain the associated semantics of the trigger words and the central sentence in the event instance, associate the associated semantics of the associated roles in the event pattern and the associated semantics of the trigger words and the central sentence in the event instance, and construct a role focus heterogeneous graph.

[0018] See also Figure 2 In step 2, the trigger word and the central sentence in the event instance are associated, the associated semantics of the trigger word and the central sentence in the event instance are obtained, the associated semantics of the associated role in the event pattern and the associated semantics of the trigger word and the central sentence in the event instance are associated, and a role focus heterogeneous graph is constructed, which specifically includes the following steps: Based on the event instance, the event trigger words and event central sentences contained in the corpus are represented respectively to obtain the trigger word node set and the central sentence node set; Establish an edge between the event trigger word and the event center sentence to obtain the associated semantics between the event trigger word and the event center sentence; Establish an edge between the event trigger word and the first event type to obtain the associated semantics between the first event trigger word and the event type; Establish an edge between the event trigger word and the first role node to obtain the associated semantics between the event trigger word and the first role node; Establish an edge between the event center sentence and the first event type to obtain the associated semantics between the event center sentence and the first event type; Establish an edge between the event center sentence and the first role node to obtain the associated semantics between the event center sentence and the first role node; According to the association semantics between the first event type and the second event type, the association semantics between the first event type and the first role node, the association semantics between the first role node and the second role node, the association semantics between the event trigger word and the event central sentence, the association semantics between the first event trigger word and the thing type, the association semantics between the event trigger word and the first role node, the association semantics between the event central sentence and the first thing type, and the association semantics between the event central sentence and the first role node, the event pattern is associated with the event instance to form a role focus heterogeneous graph.

[0019] In the step of respectively representing the event trigger words and the event central sentences contained in the corpus to obtain the trigger word node set and the central sentence node set, the corresponding process has the following relational expression: ; in, Represents a set of event trigger words, Indicates the first event trigger word, Indicates Event trigger words, Indicates the number of trigger word nodes, represents the set of central sentence nodes, Indicates Central sentence nodes, Indicates the number of central sentences, Indicates the first word of the central sentence node, Indicates the central sentence node words, Indicates the number of words; In the step of establishing an edge between the event trigger word and the event center sentence to obtain the associated semantics between the event trigger word and the event center sentence, the relationship between the corresponding process is as follows: ; In the step of establishing an edge between the event trigger word and the first event type to obtain the associated semantics between the first event trigger word and the event type, the relationship between the corresponding process is as follows: ; in, Indicates the event type corresponding to the trigger word node; In the step of establishing an edge between the event trigger word and the first role node to obtain the associated semantics between the event trigger word and the first role node, the relationship between the corresponding process is as follows: ; In the step of establishing an edge between the event center sentence and the first event type to obtain the associated semantics between the event center sentence and the first event type, the relationship between the corresponding process is as follows: ; in, Indicates the event type of the event contained in the central sentence node; In the step of establishing an edge between the event center sentence and the first role node to obtain the associated semantics between the event center sentence and the first role node, the relationship between the corresponding process is as follows: .

[0020] Furthermore, we first establish a trigger word-central sentence edge, connecting the trigger word node of each event in the corpus with the central sentence node where it is located. Both the trigger word and the central sentence contain semantic information of the event instance to a certain extent. Through this type of edge, the key role of the trigger word in the central sentence can be enhanced; Then, we establish the pattern-instance association between the event pattern level and the event instance level to form a role focus heterogeneous graph, so as to achieve a deeper understanding of the role's semantic information around the event structure in combination with the event instance. The specific steps are as follows: Establish a trigger word-event type edge, connecting the trigger word node of each event in the corpus with the event type node of the event it triggers. The same type of event is triggered by multiple trigger words of the same type, and different types of events are usually triggered by different types of trigger words. Therefore, this type of edge can transfer the semantic information of the trigger word node to the event type node, thereby indirectly affecting the semantics of the role node; Establish trigger word-role edges, connecting each trigger word node in the corpus with all role nodes of the event type to which it belongs. This type of edge can directly associate the trigger word semantics with the role, directly affecting the semantics of the role; Establish a central sentence-event type edge, and connect each central sentence node in the corpus with the event type node it contains. Generally, the central sentence contains the core / main information of the event, which can be used as the event context to supplement the instance information of the event type, and indirectly enrich the semantics of the role; Establish a central sentence-role edge, connecting each central sentence node in the corpus with the role node of its corresponding event. This type of edge can be used to associate event context information with the role, and directly enrich the semantics of the role corresponding to the event from the event instance.

[0021] Step 3: Input the role focus heterogeneous graph into the graph attention network for learning, and obtain the node representation after learning by the graph attention network.

[0022] In step 3, the role focus heterogeneous graph is input into the graph attention network for learning, and the node representation after learning by the graph attention network is obtained. The relationship between the corresponding process is as follows: ; in, Represents the graph attention network The learned node representation is express The initial representation of the nodes of the graph, express The adjacency matrix corresponding to the graph is, Indicates Type of role Representations after learning with heterogeneous graph networks.

[0023] Step 4: Use the pre-trained language model to capture document terms from the node representation learned by the graph attention network, and perform weighted fusion with the node representation learned by the graph attention network to obtain the fused semantic representation.

[0024] In step 4, the pre-trained language model is used to capture document terms on the node representation learned by the graph attention network. The relationship between the corresponding process is as follows: ; in, Represents a document word representation set, Indicates the expression of the first word, Indicates Words The expression, Indicates the document Words The expression, Indicates the first type of role. Indicates The representation of a role, Represents encoding and decoding operations, Indicates that only decoding operations are performed. Indicates the prompt. express Encoder The output result of In the step of weighted fusion of node representations learned by the graph attention network to obtain the fused semantic representation, the corresponding process has the following relationship: ; in, Indicates Type of role The fused representation of .

[0025] It should be noted that the role representation obtained by learning the role-focused heterogeneous graph network is fused with the role representation in Prompt obtained based on the pre-trained language model to further enrich the semantics of the role.

[0026] Step 5: Construct an event argument extraction model based on the role focus heterogeneous graph, and use the event argument extraction model combined with the fused semantic representation to perform cross-role argument extraction to obtain the probability of the role's span start position on the document and the probability of the role's span end position on the document; The loss function of the model is constructed based on the probability of the character's span starting position on the document and the probability of the character's span ending position on the document; The event argument extraction model is optimized based on the loss function of the model to obtain the optimized event argument extraction model.

[0027] In step 5, an event argument extraction model is constructed based on the role focus heterogeneous graph, and the event argument extraction model is combined with the fused semantic representation to perform cross-role argument extraction, so as to obtain the probability of the role's span start position on the document and the probability of the role's span end position on the document, which specifically includes the following steps: Perform span representation on the fused semantic representation to obtain a span start representation and a span end representation of the character; The event argument extraction model is used to predict the positions of the span start representation and the span end representation of the role, so as to obtain the distribution of the span start position and the span end position of the role in the document; Probability calculations are performed on the distribution of the span start position of the character on the document and the distribution of the span end position of the character on the document to obtain the probability of the span start position of the character on the document and the probability of the span end position of the character on the document.

[0028] In the step of performing span representation on the fused semantic representation to obtain the span start representation of the character and the span end representation of the character, the relationship between the corresponding processes is as follows: ; in, The span that indicates the role begins to indicate, The span that indicates the role ends. represents the starting vector, Indicates the end vector; In the step of using the event argument extraction model to predict the position of the span start representation of the role and the span end representation of the role to obtain the distribution of the span start position of the role on the document and the distribution of the span end position of the role on the document, the corresponding process has the following relationship: ; in, Representing roles the distribution of span starts over the documents, Representing roles Distribution of span endings over the document; In the step of respectively calculating the probability of the distribution of the span start position of the character on the document and the distribution of the span end position of the character on the document to obtain the probability of the span start position of the character on the document and the probability of the span end position of the character on the document, the relationship between the corresponding processes is as follows: ; in, represents the distribution of span start positions of roles on the document, represents the distribution of the span end positions of the role on the document, It means that it has been normalized; In the step of constructing the loss function of the model based on the probability of the role's span start position on the document and the probability of the role's span end position on the document, the corresponding process has the following relationship: ; in, represents the model loss function, Indicates the number of documents contained in the corpus, Indicates the documents contained in the corpus. represents the logarithmic function, Indicates the number of documents. Indicates the number of roles involved in the event.

[0029] Step 6: Use the optimized event argument extraction model to obtain the event argument results.

[0030] In step 6, argument identification is used F 1-Value and Argument Classification F 1 value is used as the evaluation indicator. If the span (i.e., the offset of the start and end positions in the document) and event type of an event argument match the span and event type mentioned by any argument, the event argument is correctly identified; if the role of the event argument is also correct, the event argument is correctly classified. The corresponding process has the following relationship: ; in, express value, Indicates the accuracy, represents the recall rate, represents the number of true predictions. represents the number of false predictions. Represents the number of false predictions that turned out to be true.

[0031] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0032] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0033] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A document-level event argument extraction method based on role-focus heterogeneous graph, characterized in that: The method comprises the following steps: Step 1: associate the event type and the role node to obtain the association semantics of the associated role in the event pattern; Step 2: Associating the trigger words and the central sentence in the event instance, obtaining the associated semantics of the trigger words and the central sentence in the event instance, associating the associated semantics of the associated roles in the event pattern and the associated semantics of the trigger words and the central sentence in the event instance, and constructing a role focus heterogeneous graph; Step 3: Input the role focus heterogeneous graph into the graph attention network for learning, and obtain the node representation after learning by the graph attention network; Step 4: Use the pre-trained language model to capture document terms from the node representation learned by the graph attention network, and perform weighted fusion with the node representation learned by the graph attention network to obtain the fused semantic representation; Step 5: Construct an event argument extraction model based on the role focus heterogeneous graph, and use the event argument extraction model combined with the fused semantic representation to perform cross-role argument extraction to obtain the probability of the role's span start position on the document and the probability of the role's span end position on the document; The loss function of the model is constructed based on the probability of the character's span starting position on the document and the probability of the character's span ending position on the document; Optimize the event argument extraction model based on the loss function of the model to obtain the optimized event argument extraction model; Step 6: Use the optimized event argument extraction model to obtain the event argument results.

2. The document-level event argument extraction method based on role-focus heterogeneous graph according to claim 1 is characterized in that: In step 1, the event type and the role node are associated to obtain the associated semantics of the associated role in the event pattern, which specifically includes the following steps: Based on the event pattern, the event type and role nodes are represented by sets, and the event type set and role set are obtained respectively; Establish an edge between the first event type and the second event type to obtain the association semantics between the first event type and the second event type; Establish an edge between the first event type and the first role node to obtain the association semantics between the first event type and the first role node; An edge is established between the first role node and the second role node to obtain the association semantics between the first role node and the second role node.

3. The document-level event argument extraction method based on role-focus heterogeneous graph according to claim 2 is characterized in that: In the steps of respectively expressing the event type and the role node as a set and obtaining the set of event types and the set of roles, the corresponding processes have the following relational expressions: ； in, Represents a set of event types. Indicates the first event type, The node number indicating the event type. Indicates the number of event types contained in the document. Indicates Event Type A collection of characters included, Indicates the first role node, Indicates The event type contains role nodes, Indicates the number of roles included in the event type; In the step of establishing an edge between the first event type and the second event type to obtain the association semantics between the first event type and the second event type, the relationship between the corresponding process is as follows: ； in, represents the adjacency matrix corresponding to the role-focus heterogeneous graph, represents the row or column of the corresponding node in the adjacency matrix, Indicates Event Type The character collection includes: Indicates Event Type Included character collection; In the step of establishing an edge between the first event type and the first role node to obtain the association semantics between the first event type and the first role node, the relationship between the corresponding process is as follows: ； in, Indicates role nodes; In the step of establishing an edge between the first role node and the second role node to obtain the association semantics between the first role node and the second role node, the relationship between the corresponding process is as follows: ； in, Indicates A role node.

4. The document-level event argument extraction method based on role-focus heterogeneous graph according to claim 3 is characterized in that: In step 2, the trigger word and the central sentence in the event instance are associated, the associated semantics of the trigger word and the central sentence in the event instance are obtained, the associated semantics of the associated role in the event pattern and the associated semantics of the trigger word and the central sentence in the event instance are associated, and a role focus heterogeneous graph is constructed, which specifically includes the following steps: Based on the event instance, the event trigger words and event central sentences contained in the corpus are represented respectively to obtain the trigger word node set and the central sentence node set; Establish an edge between the event trigger word and the event center sentence to obtain the associated semantics between the event trigger word and the event center sentence; Establish an edge between the event trigger word and the first event type to obtain the associated semantics between the first event trigger word and the event type; Establish an edge between the event trigger word and the first role node to obtain the associated semantics between the event trigger word and the first role node; Establish an edge between the event center sentence and the first event type to obtain the associated semantics between the event center sentence and the first event type; Establish an edge between the event center sentence and the first role node to obtain the associated semantics between the event center sentence and the first role node; According to the association semantics between the first event type and the second event type, the association semantics between the first event type and the first role node, the association semantics between the first role node and the second role node, the association semantics between the event trigger word and the event central sentence, the association semantics between the first event trigger word and the thing type, the association semantics between the event trigger word and the first role node, the association semantics between the event central sentence and the first thing type, and the association semantics between the event central sentence and the first role node, the event pattern is associated with the event instance to form a role focus heterogeneous graph.

5. The document-level event argument extraction method based on role-focus heterogeneous graph according to claim 4 is characterized in that: In the step of respectively representing the event trigger words and the event central sentences contained in the corpus to obtain the trigger word node set and the central sentence node set, the corresponding process has the following relational expression: ； in, Represents a set of event trigger words, Indicates the first event trigger word, Indicates Event trigger words, Indicates the number of trigger word nodes, represents the set of central sentence nodes, Indicates Central sentence nodes, Indicates the number of central sentences, Indicates the first word of the central sentence node, Indicates the central sentence node words, Indicates the number of words; In the step of establishing an edge between the event trigger word and the event center sentence to obtain the associated semantics between the event trigger word and the event center sentence, the relationship between the corresponding process is as follows: ； In the step of establishing an edge between the event trigger word and the first event type to obtain the associated semantics between the first event trigger word and the event type, the relationship between the corresponding process is as follows: ； in, Indicates the event type corresponding to the trigger word node; In the step of establishing an edge between the event trigger word and the first role node to obtain the associated semantics between the event trigger word and the first role node, the relationship between the corresponding process is as follows: ； In the step of establishing an edge between the event center sentence and the first event type to obtain the associated semantics between the event center sentence and the first event type, the relationship between the corresponding process is as follows: ； in, Indicates the event type of the event contained in the central sentence node; In the step of establishing an edge between the event center sentence and the first role node to obtain the associated semantics between the event center sentence and the first role node, the relationship between the corresponding process is as follows: 。 6. The document-level event argument extraction method based on role-focus heterogeneous graph according to claim 5 is characterized in that: In step 3, the role focus heterogeneous graph is input into the graph attention network for learning, and the node representation after learning by the graph attention network is obtained. The relationship between the corresponding process is as follows: ； in, Represents the graph attention network The learned node representation is express The initial representation of the nodes of the graph, express The adjacency matrix corresponding to the graph is, Indicates Type of role Representations after learning with heterogeneous graph networks.

7. The document-level event argument extraction method based on role-focus heterogeneous graph according to claim 6 is characterized in that: In step 4, the pre-trained language model is used to capture document terms on the node representation learned by the graph attention network. The relationship between the corresponding process is as follows: ； in, Represents a document word representation set, Indicates the expression of the first word, Indicates Words The expression, Indicates the document Words The expression, Indicates the first type of role. Indicates The representation of a role, Represents encoding and decoding operations, Indicates that only decoding operations are performed. Indicates the prompt. express Encoder The output result of In the step of weighted fusion of node representations learned by the graph attention network to obtain the fused semantic representation, the corresponding process has the following relationship: ； in, Indicates Type of role The fused representation of .

8. The document-level event argument extraction method based on role-focus heterogeneous graph according to claim 7 is characterized in that: In step 5, an event argument extraction model is constructed based on the role focus heterogeneous graph, and the event argument extraction model is combined with the fused semantic representation to perform cross-role argument extraction, so as to obtain the probability of the role's span start position on the document and the probability of the role's span end position on the document, which specifically includes the following steps: Perform span representation on the fused semantic representation to obtain a span start representation and a span end representation of the character; The event argument extraction model is used to predict the positions of the span start representation and the span end representation of the role, so as to obtain the distribution of the span start position and the span end position of the role in the document; Probability calculations are performed on the distribution of the span start position of the character on the document and the distribution of the span end position of the character on the document to obtain the probability of the span start position of the character on the document and the probability of the span end position of the character on the document.

9. The document-level event argument extraction method based on role-focus heterogeneous graph according to claim 8 is characterized in that: In the step of performing span representation on the fused semantic representation to obtain the span start representation of the character and the span end representation of the character, the relationship between the corresponding processes is as follows: ； in, The span that indicates the role begins to indicate, The span that indicates the role ends. represents the starting vector, Indicates the end vector; In the step of using the event argument extraction model to predict the position of the span start representation of the role and the span end representation of the role to obtain the distribution of the span start position of the role on the document and the distribution of the span end position of the role on the document, the corresponding process has the following relationship: ； in, Representing roles the distribution of span starts over the documents, Representing roles Distribution of span endings over the document; In the step of respectively calculating the probability of the distribution of the span start position of the character on the document and the distribution of the span end position of the character on the document to obtain the probability of the span start position of the character on the document and the probability of the span end position of the character on the document, the relationship between the corresponding processes is as follows: ； in, represents the distribution of span start positions of roles on the document, represents the distribution of span end positions of roles on the document, It means that it has been normalized; In the step of constructing the loss function of the model based on the probability of the character's span start position on the document and the probability of the character's span end position on the document, the corresponding process has the following relationship: ； in, represents the model loss function, Indicates the number of documents contained in the corpus, Indicates the documents contained in the corpus. represents the logarithmic function, Indicates the number of documents. Indicates the number of roles involved in the event.

Images