Method and system for joint extraction of events and event relations based on dynamic graph propagation

Abstract

The present invention discloses a method and system for jointly extracting events and event relationships based on dynamic graph propagation. The present invention constructs a multi-level dynamic graph structure and introduces a gate mechanism to realize the dynamic transmission and joint update of information between event internal elements and events, thereby realizing end-to-end collaborative learning of subtasks such as event trigger word recognition, argument recognition, role classification, event type determination and event relationship classification. The present invention simultaneously completes the prediction of event elements, event types and event relationships through dynamic graph propagation, effectively captures contextual dependencies, and has significant advantages in processing nested events and long-distance event relationships. It effectively improves the overall robustness and structural modeling capabilities of the event extraction system and can be widely used in financial public opinion analysis, medical event tracking, judicial case reasoning and other fields.

Description

Technical Field

[0001] The present invention belongs to the field of audio processing and natural language processing, and specifically relates to a method and system for jointly extracting events and event relationships based on dynamic graph propagation. Background Art

[0002] In the field of Natural Language Processing (NLP), event extraction is a core task in information extraction. It aims to identify semantically meaningful events and their components from unstructured text. Event extraction is widely used in scenarios such as public opinion analysis, intelligent question answering, financial information analysis, and legal text mining.

[0003] Traditional event extraction tasks typically rely on predefined event type templates or pipeline structures, following the sequence of trigger word identification, argument extraction, and role identification. However, these methods suffer from semantic fragmentation, error cascades, and insufficient structural modeling, making them difficult to handle complex, ambiguous, and nested natural language scenarios.

[0004] Furthermore, the span mechanism is widely used in current event extraction research. This mechanism uses character or word sequences of a certain length in a text as candidate entities or component units for modeling trigger words or arguments. Compared to word-level annotation, span-based modeling is more flexible and can capture complex components composed of multiple words.

[0005] To better represent the internal structure of events and the connections between them, graph neural networks (GNNs) and dynamic graph modeling methods have been increasingly applied to event extraction tasks in recent years. By constructing event feature graphs and event-level graphs, syntactic structure, semantic connections, and contextual dependencies can be more effectively captured. However, existing graph-based methods often suffer from the following issues:

[0006] 1) Static graph structures limit the ability to dynamically evolve semantics between nodes;

[0007] 2) The lack of a fine-grained graph update mechanism makes it difficult to fully integrate information from different types of nodes (such as trigger words and arguments);

[0008] 3) In the context of multiple events, there is a lack of the ability to effectively model semantic relationships between events (such as causality, parallelism, and inclusion).

[0009] Therefore, how to improve the overall robustness and structural modeling capabilities of the event extraction system is a technical problem that needs to be solved urgently. Summary of the Invention

[0010] The purpose of the present invention is to solve the problem in the prior art that events and event relations are difficult to be extracted efficiently and robustly, and to provide a method and system for jointly extracting events and event relations based on dynamic graph propagation.

[0011] The specific technical solutions adopted in the present invention are as follows:

[0012] In a first aspect, the present invention provides a method for jointly extracting events and event relations based on dynamic graph propagation, which comprises:

[0013] S1, processing the text to be extracted into a token sequence and a token embedding sequence;

[0014] S2. extracting a set of candidate spans from the token sequence, and generating an initial span representation for each candidate span by combining token embedding and span features;

[0015] S3. Based on the token part of speech, the candidate spans in the candidate span set are divided into two categories: trigger words and arguments, and the event element dynamic graph is established as the initial nodes. The final span representation of each node is obtained after updating through the gate mechanism;

[0016] S4. Input the final span representation of each node in the event element dynamic graph into the first classification model to obtain the binary classification probability of each node belonging to the trigger word and the argument, thereby updating the node type in the event element dynamic graph; then, based on the event element dynamic graph after the node type update, concatenate the final span representation of each pair of trigger word nodes and argument nodes and input them into the second classification model to obtain the event role type and confidence level corresponding to this pair of nodes, and only retain the edge connections between the trigger word nodes and the argument nodes whose confidence level is higher than the first threshold, thereby updating the graph structure;

[0017] S5. For the event element dynamic graph after the graph structure is updated, each trigger word node and all its associated argument nodes constitute an event. The final span representations of all nodes in the event are fused into the initial event representation. A fully connected event-level dynamic graph is constructed with all events as nodes. The final event representation of each event node is then obtained after updating through the gate mechanism.

[0018] S6. Input the final event representation of each event node in the event-level dynamic graph into the third classification model to predict the event type; then, concatenate the final event representations of each pair of event nodes and input them into the fourth classification model to obtain the event relationship type and confidence level, delete the edge connections between nodes with confidence levels lower than the second threshold, and form a graph of events and event relationships.

[0019] As a preferred embodiment of the above-mentioned first aspect, in S2, sliding extraction is performed on the token sequence using sliding windows of different lengths, and all possible candidate spans are extracted in an exhaustive form and added to the candidate span set, each candidate span in the candidate span set is composed of one or more consecutive tokens; then the candidate span set is screened according to syntactic dependency rules, and candidate spans with subject-predicate structure, verb-object structure, preposition-object structure or attributive-predicate structure are retained, and other candidate spans are deleted.

[0020] As a preferred embodiment of the first aspect, the initial span representation of each candidate span is formed by concatenating the start token embedding, end token embedding, average token embedding and number of tokens in the candidate span.

[0021] As a preferred embodiment of the above-mentioned first aspect, in S3, for each candidate span in the candidate span set, if the candidate span contains a subject-verb structure or a token of a verb part of speech, the candidate span is classified as a trigger word, otherwise the candidate span is classified as an argument; when initially constructing the event element dynamic graph, each trigger word node needs to establish an edge connection with all argument nodes.

[0022] As a preferred embodiment of the first aspect mentioned above, for event element dynamic graphs and event-level dynamic graphs, the gate mechanism is used to update the graph by traversing each node in the graph, calculating the weight through the gating function, and then updating the node representation of the neighboring node to the current traversed node according to the weight to complete one round of iteration; after multiple rounds of iteration, the final node representation of each node is obtained.

[0023] As a preferred embodiment of the above-mentioned first aspect, the first classification model, the second classification model, the third classification model and the fourth classification model all adopt a forward propagation neural network, and the S1~S6 constitute a joint extraction framework of events and event relationships. All learnable network parameters in the framework need to be jointly optimized using a multi-task loss function, and the multi-task loss function is obtained by weighting the trigger word prediction loss, argument recognition loss, event role classification loss, event type classification loss and event relationship classification loss.

[0024] In a second aspect, the present invention provides a system for jointly extracting events and event relationships based on dynamic graph propagation, which includes:

[0025] Tokenization module, used to process the text to be extracted into token sequences and token embedding sequences;

[0026] a span extraction module, configured to extract a set of candidate spans from the token sequence and generate an initial span representation for each candidate span by combining token embeddings and span features;

[0027] The event element dynamic graph construction module is used to classify the candidate spans in the candidate span set into trigger words and arguments based on the token part of speech, and use them as the initial nodes to build the event element dynamic graph. The final span representation of each node is obtained after updating through the gate mechanism;

[0028] The feature-level extraction module is used to input the final span representation of each node in the event feature dynamic graph into the first classification model to obtain the binary classification probability of each node belonging to the trigger word and argument, thereby updating the node type in the event feature dynamic graph. Based on the event feature dynamic graph with updated node types, the final span representation of each pair of trigger word nodes and argument nodes is concatenated and input into the second classification model to obtain the event role type and confidence level corresponding to this pair of nodes. Only the edge connections between the trigger word nodes and argument nodes with confidence levels above the first threshold are retained, thereby updating the graph structure.

[0029] The event-level dynamic graph construction module is used to construct the event element dynamic graph after the graph structure is updated. Each trigger word node and all its associated argument nodes constitute an event. The final span representation of all nodes in the event is fused into the initial event representation. A fully connected event-level dynamic graph is constructed with all events as nodes. The final event representation of each event node is then obtained after updating through the gate mechanism.

[0030] The event-level extraction module inputs the final event representation of each event node in the event-level dynamic graph into the third classification model to predict the event type; then the final event representation of each pair of event nodes is spliced ​​and input into the fourth classification model to obtain the event relationship type and confidence, and deletes the edge connections between nodes with confidence lower than the second threshold to form an event and event relationship map.

[0031] In a third aspect, the present invention provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, can implement the method for jointly extracting events and event relationships based on dynamic graph propagation as described in any one of the solutions in the first aspect above.

[0032] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for jointly extracting events and event relationships based on dynamic graph propagation as described in any one of the schemes in the first aspect above can be implemented.

[0033] In a fifth aspect, the present invention provides a computer electronic device comprising a memory and a processor;

[0034] The memory is used to store computer programs;

[0035] The processor is configured to implement the method for jointly extracting events and event relationships based on dynamic graph propagation as described in any one of the solutions of the first aspect above when executing the computer program.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention proposes a method and system for jointly extracting events and event relationships based on dynamic graph propagation. By constructing a multi-level dynamic graph structure and introducing a gate mechanism to achieve dynamic transmission and joint updating of information between event internal elements and events, the system can achieve end-to-end collaborative learning of subtasks such as event trigger word recognition, argument recognition, role classification, event type determination, and event relationship classification. The present invention simultaneously completes the prediction of event elements, event types, and event relationships through dynamic graph propagation, effectively capturing contextual dependencies, and has significant advantages in processing nested events and long-distance event relationships. It effectively improves the overall robustness and structural modeling capabilities of the event extraction system and can be widely used in fields such as financial public opinion analysis, medical event tracking, and judicial case reasoning. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the steps of the joint extraction method of events and event relations based on dynamic graph propagation;

[0039] Figure 2 Schematic diagram for calculating trigger word prediction loss, argument recognition loss, and event role classification loss;

[0040] Figure 3 Schematic diagram for calculating event type classification loss and event relationship classification loss;

[0041] Figure 4 This is a schematic diagram of the module composition of the event and event relationship joint extraction system based on dynamic graph propagation;

[0042] Figure 5 It is a structural diagram of computer electronic equipment. DETAILED DESCRIPTION

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in the various embodiments of the present invention can be combined accordingly without conflicting with each other.

[0044] In the description of the present invention, it should be understood that the terms "first" and "second" are used solely for descriptive purposes and are not to be construed as indicating or implying relative importance or implicitly specifying the number of technical features being described. Therefore, features defined as "first" or "second" may explicitly or implicitly include at least one of such features.

[0045] The present invention provides a method for jointly extracting events and event relationships based on dynamic graph propagation. By constructing a multi-level dynamic graph structure and introducing a gate mechanism to achieve dynamic transmission and joint updating of information between event elements and events, it achieves end-to-end collaborative learning of subtasks such as event trigger word recognition, argument recognition, role classification, event type determination, and event relationship classification, effectively improving the overall robustness and structural modeling capabilities of the event extraction system. The specific implementation of the above-mentioned joint extraction method of the present invention is described below.

[0046] like Figure 1 As shown, in a preferred embodiment of the present invention, a method for jointly extracting events and event relationships based on dynamic graph propagation is provided, which includes multiple steps S1 to S6. The implementation of each step is described in detail below.

[0047] S1. Process the text to be extracted into a token sequence and a token embedding sequence.

[0048] It should be noted that the text to be extracted in the present invention refers to a natural language text from which events and event relationships are to be extracted. A token in the present invention is also called a word unit. Tokenization of the text to be extracted is a state of the art. The text to be extracted can be segmented to form a token sequence. Each token is then vectorized using a language representation model to obtain a corresponding token embedding. The language representation model can be implemented using the BERT model or the Large Language Model (LLM).

[0049] In an embodiment of the present invention, the specific steps of step S1 are as follows:

[0050] S11: First, use text segmentation tools (such as Jieba, HanLP) to segment the extracted text, and then use the pre-trained BERT model to obtain a token sequence consisting of a series of tokens and obtain the initial representation of each token. ;

[0051] S12: Initial representation of each token Encoded by the bidirectional transformer encoder of the BERT model to obtain token embedding that incorporates contextual information .

[0052] To better illustrate the specific implementation of the present invention, an exemplary financial news text is provided below as the text to be extracted. The text to be extracted is recorded as Text A. The specific content is as follows: "In March 2024, due to the sharp drop in technology stocks in Country A, the stock index of Country B fell by more than 12% within a week. Company C subsequently announced a cut in its first-quarter profit forecast and laid off 2,000 employees to control costs. This move further aroused market concerns about the health of the banking industry, and relevant financial regulatory authorities also intervened in the investigation."

[0053] The text A to be extracted is tokenized and encoded using BERT to obtain token-level context embedding, thereby obtaining a rich semantic representation. Some examples of tokens extracted are "plunge", "fall", "cut profit expectations", "layoffs", "concerns", and "involvement in investigations".

[0054] S2. Extract candidate spans from the token sequence and form a candidate span set, and generate an initial span representation of each candidate span in the candidate span set by combining token embedding and span features.

[0055] It should be noted that each span is a token sequence of a certain length, at least one token, and can also be multiple tokens.

[0056] In an embodiment of the present invention, sliding windows of varying lengths can be used to perform sliding extraction on the token sequence, exhaustively extracting all possible candidate spans and adding them to a candidate span set. Each candidate span in the candidate span set consists of one or more consecutive tokens. The candidate span set is then screened based on syntactic dependency rules, retaining those with subject-verb structures (SBV), verb-object structures (VOB), prepositional-object structures (POB), or attributive-medial structures (ATT), or those containing only a single noun or verb part-of-speech token, while removing all other candidate spans. For subject-verb structures, the dependency relationship between the subject and verb is extracted to form the candidate spans; for verb-object structures, the dependency relationship between the verb and object is extracted to form the candidate spans; for prepositional-object structures, the dependency relationship between the preposition and its object is extracted to form the candidate spans; and for attributive-medial structures, the relationship between the adjective modifying the noun and the noun is extracted to form the candidate spans. Finally, all candidate spans are screened and merged to ensure that the result is semantically complete and meets structural requirements, thus forming the candidate span set.

[0057] Continuing with the example of the text A to be extracted, in step S2, exhaustive sliding window extraction is combined with syntactic dependency structure rules to generate candidate spans. For example, the original token segmentation for "This move has triggered further market concerns about the health of the banking industry" is ["this move", "trigger", "market", "to", "banking industry", "health", "of", "further", "concern"]. Assuming the total number of tokens in text A to be extracted is n, exhaustive extraction of all candidate spans of 1 to n tokens using a sliding window of variable length is performed and added to the candidate span set. For example, the candidate spans of 2 tokens might be "this move triggered", "triggering the market", "market to", etc. The dependency structure of the candidate span set is analyzed (or simplified) using tools such as spaCy or Stanza. The candidate span set is filtered using dependency structure rules, including those that meet the subject-verb structure (SBV), verb-object structure (VOB), preposition-object structure (POB), and attributive-text structure (ATT), such as "triggered," "triggered market concerns," and "further concerns." The remaining candidate spans are deleted. For each retained candidate span, an initial span representation is generated by combining token embeddings and span features.

[0058] In an embodiment of the present invention, the initial span representation of each candidate span is formed by concatenating the start token embedding, end token embedding, average token embedding, and number of tokens in the candidate span. The calculation formula can be expressed as:

[0059]

[0060] in is the token embedding of the i-th candidate span start token, The token embedding for this candidate span end token, is the average of the token embeddings of all tokens in this candidate span, It encodes the number of tokens contained in the candidate span, and “;” indicates vector concatenation.

[0061] S3. Based on the token part of speech, the candidate spans in the candidate span set are divided into two categories: trigger words and arguments, and the event element dynamic graph is established as the initial node. The final span representation of each node is obtained after updating through the gate mechanism.

[0062] It's important to note that in event extraction, triggers, arguments, and event roles are three core concepts used to describe key information about an event. Triggers are the core indicators of an event's occurrence, identifying its type. They determine the event's category and serve as the starting point for event extraction. For example, in the sentence "Xiao Ming signed a contract in Beijing yesterday," "signed" is the trigger, indicating a "contract signing" event; in the sentence "The company announced layoffs," "announced" is the trigger, indicating an "announcement" event. Arguments, on the other hand, are participants or attributes of an event and are crucial components of an event. They provide detailed information about the event, including the subject, object, time, and location. In "Xiao Ming signed a contract in Beijing yesterday," the subject is "Xiao Ming," the time is "yesterday," the location is "Beijing," and the object is "contract." In "The company announced layoffs," the subject is "company," and the object is "layoffs." By identifying triggers and arguments, we can fully extract event information, providing a foundation for event extraction. Event roles refer to entities or concepts that assume specific functions or attributes within an event. They describe the specific role or identity of arguments within an event and are an important component of event structure. For example, in the sentence "Xiao Ming signed a contract in Beijing yesterday," if "Xiao Ming" is the argument, its event role might be "Contract Signer."

[0063] Therefore, when the candidate spans in the candidate span set are divided into trigger words and arguments, they can be distinguished based on their general parts of speech. In the present invention, for each candidate span in the candidate span set, if the candidate span contains a verb-object structure or a token with a verb part of speech, the candidate span is classified as a trigger word; otherwise, the candidate span is classified as an argument. In this way, all trigger words and arguments can be used as two-node graph nodes. When initially constructing the event element dynamic graph, each trigger word node in the initial graph needs to establish an edge connection with all argument nodes, and then the edge connections in the graph are pruned based on the confidence level.

[0064] Continuing with the example of the text to be extracted, A, when executing the event element dynamic graph in step S3, the initial trigger word nodes are selected as candidate spans with verbal parts of speech, such as "plunge," "announce," "cut," "layoff," and "investigation." The remaining spans serve as initial argument nodes. Initial edges are connected between each trigger word and all candidate arguments, with the edge type being "role to be classified." This allows for dynamic updates of node representations through information propagation mechanisms between trigger words and arguments within the graph structure (e.g., gated neural networks).

[0065] In an embodiment of the present invention, a gate mechanism is used to update the event element dynamic graph. Specifically, each node in the graph is traversed, and the weight is calculated using a gating function. Then, the node representation of the neighboring node is updated to the currently traversed node based on the weight, thereby propagating information between the trigger word and the argument, completing one round of iteration. After multiple rounds of iteration, the final node representation of each node is obtained. Specifically, the process of updating the event element dynamic graph using the gate mechanism can be performed according to the following sub-steps:

[0066] S31: Traverse each trigger word node and argument node in the event element dynamic graph. For each currently traversed node i, its current latest node representation is recorded as , you need to first express the node Generate a neighbor aggregation feature , generated as follows:

[0067] Assume that the neighbor node set of the currently traversed node i is , first traverse all neighbor nodes j of node i, and calculate the node representation of the current traversed node i and the node representation of neighbor node j The similarity between , and then normalize the similarity score to get the normalized weight between the two :

[0068]

[0069] This can be done using normalized weights Node representation for all neighboring nodes Perform weighted aggregation to obtain neighbor aggregation features = .

[0070] S32: Using Neighbor Aggregation Features , updated according to the gate mechanism , update as follows:

[0071]

[0072] in, is the current node representation, λ ‌ is the weight generated by the gating function, which is calculated using the following function:

[0073] )

[0074] Among them, σ is the simoid activation function, is the trainable parameter matrix, [ Indicates that the current node representation is concatenated with the updated value.

[0075] S33: Repeat steps S31 and S32 N times to obtain the final node representation of the currently traversed node i Since each node is a candidate span, the final node represents This is the final span representation.

[0076] S4. Input the final span representation of each node in the event element dynamic graph into the first classification model to obtain the binary classification probability of each node belonging to the trigger word and argument, thereby updating the node type in the event element dynamic graph; then based on the event element dynamic graph after the node type is updated, splice the final span representation of each pair of trigger word nodes and argument nodes and input them into the second classification model to obtain the event role type and confidence corresponding to this pair of nodes, and only retain the edge connections between the trigger word nodes and argument nodes whose confidence is higher than the first threshold, thereby updating the graph structure.

[0077] It should be noted that the first classification model and the second classification model in the present invention can adopt any model that can realize classification, such as a multi-layer perceptron MLP, a forward propagation neural network FNN, a support vector machine SVM, etc. In the embodiment of the present invention, the first classification model and the second classification model can both be realized by a forward propagation neural network FNN, and thus the specific implementation sub-steps of the above-mentioned step S4 are as follows:

[0078] S41: Final node representation of each node in the event element dynamic graph through the first feed-forward neural network (FNN) Classify and predict the binary probability of it being a trigger word or argument, thereby updating the node type in the event element dynamic graph.

[0079] S42: For each pair of neighbor nodes in the event element dynamic graph, the final node of the two nodes is represented as The splicing is performed and input into the second forward propagation neural network to predict the event role type and probability between the node pair.

[0080] S43: The probability of the event role type predicted for each pair of neighbor nodes in the event element dynamic graph is used as the confidence, and the event element graph structure is updated according to the confidence. Only the edge connections between the trigger word nodes and the argument nodes with confidence greater than the preset threshold are retained, and the rest of the edge connections need to be removed, thereby completing the dynamic adjustment of the event element dynamic graph structure.

[0081] Continuing with the above text A to be extracted as an example, after retaining the edges with confidence higher than the threshold of 0.8, examples of event elements such as trigger words, arguments, and event roles retained in the event element dynamic graph are shown in Table 1:

[0082] Table 1

[0083] trigger words Argument Event Role Confidence Plummet Technology stocks in Country A Affected Objects 0.91 Fall B Stock Index Affected Objects 0.88 reduce Profit expectations Operation Object 0.85 layoffs 2,000 people The recipient 0.93 investigation Financial regulators Executor 0.90

[0084] S5. For the event element dynamic graph after the graph structure is updated, each trigger word node and all its associated argument nodes constitute an event. The final span representation of all nodes in the event is fused into the initial event representation, and a fully connected event-level dynamic graph is constructed with all events as nodes. The final event representation of each event node is obtained after updating through the gate mechanism. .

[0085] It's important to note that a trigger word node in an event element dynamic graph may be connected to multiple argument nodes. Each trigger word node and all its associated argument nodes constitute an event, allowing a series of events to be extracted from the event element dynamic graph. These events can be used as nodes to construct an event-level dynamic graph. The initial event-level dynamic graph is a fully connected graph, meaning that any two event nodes have an initial edge connection, but the edge type is "unclassified event relationship."

[0086] Continuing with the above text A to be extracted as an example, the event elements shown in Table 1 are integrated to construct five structured event examples as shown in Table 2:

[0087] Table 2

[0088] Event Number event elements E1 Plummet (trigger: "plunge", argument: "country A tech stocks") E2 Index down (trigger: "down", argument: "B stock index") E3 Earnings forecast adjustment (trigger: "cut", argument: "earnings forecast") E4 Layoffs (trigger: "layoffs", argument: "2000 people") E5 Regulatory investigation (trigger: "investigation", argument: "financial regulator")

[0089] The gate mechanism can also be used to update the event-level dynamic graph. The update method is the same as the gate mechanism update method of the event element dynamic graph. Each event node in the event-level dynamic graph can be traversed, and the weight between the current traversed node and each of its neighbor nodes is calculated through the gate function. Then, the node representation of the neighbor node is updated to the current traversed node according to their respective weights to complete a round of iteration. After multiple rounds of iteration, each event node is obtained. k The final event representation Therefore, by introducing a gate mechanism into the event-level dynamic graph, referring to the information propagation process of steps S31 to S33, information is transmitted between event nodes and event representations are updated, thereby improving the modeling capabilities of event types and event relationships.

[0090] S6. Input the final event representation of each event node in the event-level dynamic graph into the third classification model to predict the event type; then, concatenate the final event representations of each pair of event nodes and input them into the fourth classification model to obtain the event relationship type and confidence level, delete the edge connections between nodes with confidence levels lower than the second threshold, and form a graph of events and event relationships.

[0091] It should be noted that the third classification model and the fourth classification model in the present invention can adopt any model that can implement classification, such as a multi-layer perceptron MLP, a forward propagation neural network FNN, a support vector machine SVM, etc. In the embodiment of the present invention, the third classification model and the fourth classification model can both be implemented using a forward propagation neural network FNN, and thus the specific implementation sub-steps of the above-mentioned step S6 are as follows:

[0092] S61: Represent the final event of each event node in the event-level dynamic graph , classification is performed through a forward propagation network to predict the event type and confidence level corresponding to the event node.

[0093] Continuing with the above text A to be extracted as an example, for the five events shown in Table 2 above, the remaining event type prediction results are shown in Table 3 below:

[0094] Table 3

[0095] Event Number Event Type E1 Stock market change events E2 Stock market change events E3 Financial behavior E4 Personnel changes E5 Policy response

[0096] S62: Final event representation for each pair of neighbor nodes in the event-level dynamic graph After splicing, another forward propagation network is used to predict the event relationship type and confidence. The event relationship label can include sequential relationship, causal relationship, inclusion relationship, triggering relationship, etc.

[0097] Continuing with the above text A to be extracted as an example, for the five events shown in Table 2 above, the event relationship type prediction results are shown in Table 4 below:

[0098] Table 4

[0099]

[0100] S63: After predicting the event relationship type and confidence for each pair of neighbor nodes in the event-level dynamic graph, retain the high-confidence nodes and edges, and delete the edge connections between nodes with confidence lower than a preset threshold, thereby forming the final event and event relationship graph.

[0101] It should be noted that the event and event relationship joint extraction method based on dynamic graph propagation described in S1 to S6 above can constitute a joint extraction framework for events and event relationships as a whole. All learnable network parameters in the framework need to be jointly optimized using a multi-task loss function, and the first classification model, the second classification model, the third classification model, the fourth classification model, and the learnable parameters in other modules need to participate in the optimization. The above multi-task loss function is obtained by weighting the trigger word prediction loss, the argument recognition loss, the event role classification loss, the event type classification loss, and the event relationship classification loss, where Figure 2As shown in , the trigger word prediction loss, argument recognition loss, and event role classification loss are calculated by the first classification model and the second classification model based on the event element dynamic graph, and as shown in Figure 3 As shown, the event type classification loss and event relationship classification loss are calculated by the third classification model and the fourth classification model based on the event-level dynamic graph. The specific model training method belongs to the prior art. In an embodiment of the present invention, the network training steps include:

[0102] S71: Construct a training dataset containing event trigger words, argument roles, event types, and event relationship annotations;

[0103] S72: Sampling training samples from the training data set, executing steps S1 to S6 for each training sample, and calculating the predicted probabilities of trigger words, arguments, roles, event types, and event relationships;

[0104] S73: Use a multi-task loss function to jointly optimize each subtask, where the loss function includes:

[0105] Trigger word prediction loss

[0106] Argument identification loss

[0107] Role classification loss

[0108] Event type classification loss

[0109] Event relation classification loss

[0110] The joint loss function is defined as:

[0111] + + + +

[0112] in There are five adjustable weight coefficients used to balance the importance of each subtask.

[0113] S74: Based on the training data set and the joint loss function, the network parameters are iteratively updated using the back propagation algorithm and the gradient descent strategy until the loss convergence.

[0114] As a result, the aforementioned event and event relationship graph records the ultimately retained events, event types, and event relationships. Each event also records the corresponding event elements (trigger words, arguments, and event roles). It should be noted that the event and event relationship joint extraction results ultimately returned to the user can be a visualization of the aforementioned event element dynamic graph, event and event relationship graph, or other user-specified formats, such as structured JSON representation. When displaying, the event and event relationship joint extraction results are preferably returned in the form of a visual graph, and when storing, they are preferably stored as a structured JSON file.

[0115] It should be noted that, each of the method steps shown in S1 to S6 above can essentially be implemented in the form of a computer program.

[0116] Therefore, based on the same inventive concept, a joint extraction system of events and event relations based on dynamic graph propagation is also provided, such as Figure 4 As shown, the system includes:

[0117] Tokenization module, used to process the text to be extracted into token sequences and token embedding sequences;

[0118] a span extraction module, configured to extract a set of candidate spans from the token sequence and generate an initial span representation for each candidate span by combining token embeddings and span features;

[0119] The event element dynamic graph construction module is used to classify the candidate spans in the candidate span set into trigger words and arguments based on the token part of speech, and use them as the initial nodes to build the event element dynamic graph. The final span representation of each node is obtained after updating through the gate mechanism;

[0120] The feature-level extraction module is used to input the final span representation of each node in the event feature dynamic graph into the first classification model to obtain the binary classification probability of each node belonging to the trigger word and argument, thereby updating the node type in the event feature dynamic graph. Based on the event feature dynamic graph with updated node types, the final span representation of each pair of trigger word nodes and argument nodes is concatenated and input into the second classification model to obtain the event role type and confidence level corresponding to this pair of nodes. Only the edge connections between the trigger word nodes and argument nodes with confidence levels above the first threshold are retained, thereby updating the graph structure.

[0121] The event-level dynamic graph construction module is used to construct the event element dynamic graph after the graph structure is updated. Each trigger word node and all its associated argument nodes constitute an event. The final span representation of all nodes in the event is fused into the initial event representation. A fully connected event-level dynamic graph is constructed with all events as nodes. The final event representation of each event node is then obtained after updating through the gate mechanism.

[0122] The event-level extraction module inputs the final event representation of each event node in the event-level dynamic graph into the third classification model to predict the event type; then the final event representation of each pair of event nodes is spliced ​​and input into the fourth classification model to obtain the event relationship type and confidence, and deletes the edge connections between nodes with confidence lower than the second threshold to form an event and event relationship map.

[0123] In addition, based on the same inventive concept, Figure 5 As shown, the present invention also provides a computer electronic device corresponding to the method for jointly extracting events and event relations based on dynamic graph propagation provided in the above embodiment, which includes a memory and a processor;

[0124] The memory is used to store computer programs;

[0125] The processor is configured to implement the aforementioned method for jointly extracting events and event relations based on dynamic graph propagation when executing the computer program;

[0126] Furthermore, the logic instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention.

[0127] Therefore, based on the same inventive concept, the present invention provides a computer-readable storage medium corresponding to a method for jointly extracting events and event relationships based on dynamic graph propagation, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it can implement the method for jointly extracting events and event relationships based on dynamic graph propagation as described above.

[0128] Therefore, based on the same inventive concept, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the aforementioned method for joint extraction of events and event relationships based on dynamic graph propagation.

[0129] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by the processor to perform the above steps S1 to S6.

[0130] It is understood that the storage medium may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage medium may be any medium capable of storing program code, such as a USB flash drive, a mobile hard drive, a magnetic disk, or an optical disk.

[0131] It is understandable that the above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

[0132] It should also be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the various embodiments provided in this application, the division of steps or modules in the system and method is only a logical function division. In actual implementation, there may be other division methods, for example, multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0133] The embodiments described above are merely some preferred implementations of the present invention and are not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A method for jointly extracting events and event relations based on dynamic graph propagation, characterized in that: include: S1, processing the text to be extracted into a token sequence and a token embedding sequence; S2. extracting a set of candidate spans from the token sequence, and generating an initial span representation for each candidate span by combining token embedding and span features; S3. Based on the token part of speech, the candidate spans in the candidate span set are divided into two categories: trigger words and arguments, and the event element dynamic graph is established as the initial nodes. The final span representation of each node is obtained after updating through the gate mechanism; S4. Inputting the final span representation of each node in the event element dynamic graph into the first classification model to obtain the binary classification probability of each node belonging to the trigger word and the argument, thereby updating the node type in the event element dynamic graph; Based on the event element dynamic graph after the node type update, the final span representation of each pair of trigger word nodes and argument nodes is concatenated and input into the second classification model to obtain the event role type and confidence level corresponding to this pair of nodes. Only the edge connections between the trigger word nodes and argument nodes with confidence levels above the first threshold are retained, thereby updating the graph structure. S5. For the event element dynamic graph after the graph structure is updated, each trigger word node and all its associated argument nodes constitute an event. The final span representations of all nodes in the event are fused into the initial event representation. A fully connected event-level dynamic graph is constructed with all events as nodes. The final event representation of each event node is then obtained after updating through the gate mechanism. S6. Input the final event representation of each event node in the event-level dynamic graph into the third classification model to predict the event type; The final event representations of each pair of event nodes are then concatenated and input into the fourth classification model to obtain the event relationship type and confidence. The edge connections between nodes with confidence levels lower than the second threshold are deleted to form an event and event relationship graph that records the final retained events, event types, and event relationships.

2. The method for jointly extracting events and event relations based on dynamic graph propagation according to claim 1, characterized in that: In S2, sliding extraction is performed on the token sequence using sliding windows of different lengths, and all possible candidate spans are extracted in an exhaustive manner and added to a candidate span set, where each candidate span in the candidate span set consists of one or more consecutive tokens; then the candidate span set is screened according to syntactic dependency rules, and candidate spans with subject-predicate structure, verb-object structure, preposition-object structure, or attributive-predicate structure are retained, and other candidate spans are deleted.

3. The method for jointly extracting events and event relations based on dynamic graph propagation according to claim 1, characterized in that: The initial span representation of each candidate span is concatenated by the start token embedding, end token embedding, average token embedding, and number of tokens in the candidate span.

4. The method for jointly extracting events and event relations based on dynamic graph propagation according to claim 1, characterized in that: In S3, for each candidate span in the candidate span set, if the candidate span is a token of a verb part of speech, the candidate span is classified as a trigger word, otherwise the candidate span is classified as an argument; when initially constructing the event element dynamic graph, each trigger word node needs to establish an edge connection with all argument nodes.

5. The method for jointly extracting events and event relations based on dynamic graph propagation according to claim 1, characterized in that: For event element dynamic graphs and event-level dynamic graphs, the gate mechanism is used to update the graph by traversing each node in the graph, calculating the weight through the gating function, and then updating the node representation of the neighboring node to the current traversed node based on the weight to complete a round of iteration; after multiple rounds of iteration, the final node representation of each node is obtained.

6. The method for jointly extracting events and event relations based on dynamic graph propagation according to claim 1, characterized in that: The first classification model, the second classification model, the third classification model and the fourth classification model all adopt a forward propagation neural network, and S1~S6 constitute a joint extraction framework of events and event relationships. All learnable network parameters in the framework need to be jointly optimized using a multi-task loss function, and the multi-task loss function is obtained by weighting the trigger word prediction loss, argument recognition loss, event role classification loss, event type classification loss and event relationship classification loss.

7. A joint extraction system of events and event relations based on dynamic graph propagation, characterized in that: include: Tokenization module, used to process the text to be extracted into token sequences and token embedding sequences; a span extraction module, configured to extract a set of candidate spans from the token sequence and generate an initial span representation for each candidate span by combining token embeddings and span features; The event element dynamic graph construction module is used to classify the candidate spans in the candidate span set into trigger words and arguments based on the token part of speech, and use them as the initial nodes to build the event element dynamic graph. The final span representation of each node is obtained after updating through the gate mechanism; The feature-level extraction module is used to input the final span representation of each node in the event feature dynamic graph into the first classification model to obtain the binary classification probability of each node belonging to the trigger word and argument, thereby updating the node type in the event feature dynamic graph; Based on the event element dynamic graph after the node type update, the final span representation of each pair of trigger word nodes and argument nodes is concatenated and input into the second classification model to obtain the event role type and confidence level corresponding to this pair of nodes. Only the edge connections between the trigger word nodes and argument nodes with confidence levels above the first threshold are retained, thereby updating the graph structure. The event-level dynamic graph construction module is used to construct the event element dynamic graph after the graph structure is updated. Each trigger word node and all its associated argument nodes constitute an event. The final span representation of all nodes in the event is fused into the initial event representation. A fully connected event-level dynamic graph is constructed with all events as nodes. The final event representation of each event node is then obtained after updating through the gate mechanism. The event-level extraction module inputs the final event representation of each event node in the event-level dynamic graph into the third classification model to predict the event type; The final event representations of each pair of event nodes are then concatenated and input into the fourth classification model to obtain the event relationship type and confidence. The edge connections between nodes with confidence levels lower than the second threshold are deleted to form an event and event relationship graph that records the final retained events, event types, and event relationships.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, it can implement the method for jointly extracting events and event relationships based on dynamic graph propagation as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method for jointly extracting events and event relationships based on dynamic graph propagation according to any one of claims 1 to 6 is implemented.

10. A computer electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is configured to implement the method for jointly extracting events and event relationships based on dynamic graph propagation as described in any one of claims 1 to 6 when executing the computer program.

Images