Event relation extraction method and system based on dynamic graph relation reasoning

Abstract

The invention discloses an event relation extraction method and system based on dynamic graph relation reasoning, and belongs to the technical field of natural language processing. The method comprises the following steps: performing event mention recognition and semantic coding on an input text to generate a semantic representation vector; dynamically constructing a multi-relation event graph based on the vector, and learning time sequence, causality and co-reference relation weights among events through a relation perception attention mechanism; carrying out message passing and node updating by adopting a relation perception graph neural network, and explicitly modeling logic constraints among relations through a cross-relation interaction layer; multi-label relation classification is carried out based on enhanced event representation, and the same event pair is supported to have multiple relations; an external knowledge fusion mechanism is introduced to enhance reasoning reasonability, end-to-end optimization is performed by adopting a joint loss function containing a logic constraint item, and rapid migration of a low-resource scene is realized in combination with a meta-learning adaptation mechanism. The problem that a traditional method is difficult to process multi-relation co-occurrence, long-distance dependence and low-resource adaptation is solved, and the accuracy, robustness and generalization ability of event relation extraction are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing and artificial intelligence technology, specifically relating to a method and system for extracting semantic relationships between events from unstructured text. In particular, it is an event relationship extraction technique that dynamically constructs multiple relationship graphs and utilizes relationship-aware graph neural networks for joint reasoning. This method can be widely applied to scenarios requiring a deep understanding of the event logic within text, such as public opinion analysis, intelligent question answering, and knowledge graph construction. Background Technology

[0002] Event relation extraction is a core task in the field of information extraction, aiming to identify semantic relationships between event mentions in text, such as temporal relationships, causal relationships, and coreference relationships. Accurately extracting these relationships is crucial for constructing event knowledge chains and understanding narrative logic.

[0003] Existing technical solutions mainly face the following challenges: Isolated relation processing: Traditional methods often treat different types of relations (such as temporal and causal) as independent tasks, ignoring the inherent logical connections and constraints between them. For example, causal relationships usually imply temporal sequence. Limitations of static modeling: Methods based on predefined rules or static graph structures (such as dependency syntax trees) lack the ability to adapt to context and struggle to handle dynamically changing event relationships in complex contexts. Long-distance dependencies and sparsity: When event mentions are scattered throughout long documents or labeled data is scarce, the model struggles to effectively capture long-distance dependencies and is prone to declining generalization ability due to insufficient data. Conflicts in the co-occurrence of multiple relations: Real-world texts often contain pairs of events with multiple relations simultaneously; existing methods are insufficient in ensuring the logical consistency of multi-label prediction.

[0004] While existing research has attempted to partially address the aforementioned problems through joint learning, graph neural networks, or the introduction of external knowledge, breakthroughs remain in how to dynamically and explicitly model the interactions and constraints between multiple relationships in a unified manner, and effectively adapt to complex scenarios such as low resources and long documents. Therefore, a novel method for event relationship extraction is urgently needed, capable of uniformly modeling the interactions of multiple relationships, adaptively and dynamically constructing semantic associations, and possessing strong generalization capabilities. This invention, based on dynamic graph neural networks and meta-learning mechanisms, aims to solve the above problems and provide an effective solution for event relationship extraction in complex contexts. Summary of the Invention

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: In the first aspect, the present invention proposes an event extraction method based on multi-scale semantic understanding and latent attention regulation, the core technical solution of which includes four key modules: event mention recognition and semantic encoding, dynamic multi-relation graph construction, relation-aware graph neural network reasoning, and multi-label relation classification and joint optimization.

[0006] As a further improvement of the present invention, the "event mention recognition and semantic encoding" specifically includes: preprocessing the input text and identifying event mentions, and using a pre-trained language model to obtain the context-related semantic representation vector of each event mention. This step provides a high-quality semantic foundation for subsequent analysis.

[0007] As a further improvement of the present invention, the "dynamic multi-relationship graph construction" specifically includes: dynamically constructing a multi-relationship event graph and querying common sense relationship evidence related to event mentions from a structured knowledge base, and adaptively integrating external knowledge into the event representation through a gating mechanism to generate knowledge-enhanced event representations.

[0008] As a further improvement of the present invention, the "relationship-aware graph neural network reasoning" specifically includes: for the constructed dynamic multi-relation graph, a multi-head graph attention mechanism is adopted to perform message passing and node representation updates in a specific semantic space of each relationship, generating relationship-enhanced event representations. Furthermore, the mutual influence and logical constraints between different relationship types are explicitly modeled through a relationship interaction layer to improve joint reasoning capabilities; the multi-relation representations after interaction are fused to generate the final relationship-enhanced event representation.

[0009] As a further improvement of the present invention, the "multi-label relationship classification and joint optimization" specifically includes: classifying event mention pairs into multi-label relationships based on the relationship-enhanced event representation, and predicting the relationship types they exist; setting logical constraints and enhancing the constraint loss term of logical consistency; adopting end-to-end joint optimization to simultaneously optimize the event mention recognition and relationship classification tasks; introducing a meta-learning adaptation mechanism to encode the semantic features of the current task through task embedding vectors, and generating model parameter increments based on task embeddings to achieve rapid adaptation of the model in low-resource scenarios.

[0010] As a further improvement of the present invention, the association weights are dynamically calculated in the "dynamic multi-relationship graph construction" step, preferably using a mechanism based on attention coefficients to independently learn the connection strength between nodes for each relationship type.

[0011] As a further improvement of the present invention, the "dynamic multi-relationship graph construction" step introduces external knowledge such as evidence of inter-event correlation obtained from common sense knowledge base, and performs adaptive fusion with the internal representation of the model through a gating mechanism to enhance the rationality of reasoning.

[0012] As a further improvement of the present invention, the relation interaction layer in the "relation-aware graph neural network reasoning" step can realize information interaction and constraint implementation between different relation representations through cross-relation attention calculation.

[0013] As a further improvement of the present invention, the meta-learning adaptation in the "multi-label relationship classification and joint optimization" step generates task embeddings from a small number of support set samples and quickly adjusts some parameters of the model based on them, enabling the model to quickly adapt to new domains or new relationship patterns and significantly improve performance in low-resource scenarios.

[0014] This invention provides an event relationship extraction system, comprising: an event encoding module for performing the "event mention recognition and semantic encoding"; a dynamic graph construction module for performing the "dynamic multi-relationship graph construction"; a relationship reasoning module for performing the "relationship-aware graph neural network reasoning"; and a relationship classification module for performing the "multi-label relationship classification and joint optimization".

[0015] Thirdly, the present invention provides a computing device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the program to implement the method described in the first aspect.

[0016] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0017] The beneficial effects of this invention are as follows: First, by using a dynamic multi-relationship graph and a relationship interaction layer, it achieves unified and joint modeling of various event relationships, explicitly capturing the dependencies and constraints between relationships and improving the consistency of inference. Second, the dynamic graph construction mechanism enables the model to adaptively learn event associations according to the specific textual context, overcoming the rigidity problem of static structures and making it more effective in handling complex contexts and long documents. Third, the introduction of a meta-learning mechanism enables the model to quickly adapt to limited labeled data, improving its practicality in low-resource scenarios and cross-domain tasks. Finally, by introducing logical constraint loss, it effectively reduces logical conflicts in multi-label prediction and improves the reliability of the results. Attached Figure Description

[0018] Figure 1 is a flowchart of the event relationship extraction method based on dynamic graph relationship reasoning in an embodiment of the present invention.

[0019] Figure 2 is a flowchart of event mention identification and semantic encoding in the event relationship extraction method based on dynamic graph relationship reasoning in an embodiment of the present invention.

[0020] Figure 3 is a flowchart of the dynamic multi-relationship graph construction process in the event relationship extraction method based on dynamic graph relationship reasoning in an embodiment of the present invention.

[0021] Figure 4 is a flowchart of the relation-aware graph neural network reasoning process in the event relation extraction method based on dynamic graph relation reasoning in an embodiment of the present invention.

[0022] Figure 5 is a flowchart of multi-label relationship classification and joint optimization in the event relationship extraction method based on dynamic graph relationship reasoning in this embodiment of the invention.

[0023] Figure 6 is a flowchart of the adaptive fusion of external knowledge in the event relationship extraction method based on dynamic graph relationship reasoning in an embodiment of the present invention.

[0024] Figure 7 is a flowchart of meta-learning adaptation in the event relationship extraction method based on dynamic graph relationship reasoning in an embodiment of the present invention. Detailed Implementation

[0025] The event relationship extraction method of the present invention mainly includes four key implementation steps, each step being interconnected to form a complete processing pipeline. The specific implementation details of each step will be described in detail below with reference to the accompanying drawings.

[0026] As shown in Figure 1, the method includes the following steps: Step S100, event mention recognition and semantic encoding. This step is the data preprocessing and basic feature extraction stage of the method, aiming to transform the original unstructured text input into a machine-understandable, semantically rich structured numerical representation, laying a solid foundation for subsequent complex relational reasoning. Its specific implementation includes the following sub-steps, such as... Figure 2 As shown: First, the input raw text undergoes preprocessing and event mention identification. Preprocessing includes basic operations such as text standardization and cleaning, word segmentation, or sub-word partitioning to make it meet the input requirements of the downstream model. Event mention identification refers to automatically identifying and extracting all core elements representing the occurrence of events in the text through techniques such as sequence labeling, span classification, or trigger word detection. These mainly include event trigger words (such as verbs and nouns) and their related arguments (such as agent and patient). This process concretizes the event instances in the text into a series of objects to be analyzed. Second, deep semantic encoding is performed. The identified event mentions and their surrounding text fragments are input into a deep language model pre-trained on a large-scale corpus for encoding. This pre-trained language model is based on the Transformer architecture and can deeply understand the precise meaning and grammatical relationships of words in specific contexts. After encoding the input text, the model outputs a context-related semantic representation vector for each event mention. These vectors are high-dimensional, dense numerical arrays that not only capture the semantics of the event mention itself but also integrate the global contextual information of the sentence in which it is located and its dependencies with other words.

[0027] The output of this step is a fixed-dimensional semantic vector representation for each event mention. These representations form the core data foundation for all subsequent analyses, and their quality directly determines the depth and accuracy of the model's understanding of event semantics, providing high-quality semantic feature input for subsequent dynamic graph construction and relational reasoning.

[0028] As shown in Figure 1, the method includes the following steps: Step S200, dynamic multi-relation graph construction. This step is one of the core innovations of this method, aiming to organize discrete event mentions in the text into a graph structure that can dynamically reflect their potential semantic associations, providing an adaptive, relation-aware topological foundation for subsequent relational joint reasoning. The specific construction process is as follows... Figure 3 As shown: First, the graph nodes are initialized. All event mentions identified in step S100 are defined as the set of nodes in the graph, and each node is initialized using its corresponding semantic representation vector. Second, a multi-relation graph structure is defined. Unlike existing technologies that rely on fixed grammar rules or predefined templates to construct static graphs, this invention abandons this rigid structure and instead constructs a relation-specific subgraph independently for each type of semantic relation to be extracted (e.g., temporal relation, causal relation, coreference relation, etc.). This means that for the same set of event nodes, the system will maintain multiple relation views in parallel. Then, the core lies in dynamically calculating relation-specific edge weights. For any pair of event nodes, under each relation type, this invention uses a relation-aware attention mechanism to dynamically calculate the association strength between them. This mechanism first projects the node vectors into a specific feature space that matches the semantics of the relation, and then evaluates the probability of the existence of this relation by calculating the interaction score of the nodes in this space. This calculation process is entirely driven by model parameters and depends on the specific context of the input text, so that the connection weights between nodes can be adaptively learned rather than pre-set. Finally, a relation-specific adjacency matrix is ​​generated. Based on the association weights between all node pairs calculated above, a corresponding adjacency matrix is ​​generated for each relation type. This matrix accurately represents the existence and strength of a connection between any two nodes in the graph from the perspective of a specific relation.

[0029] Through this step, the present invention successfully transforms unstructured text sequences into a series of structured, relationally defined dynamic graphs. These graph structures are not fixed but can flexibly evolve according to the semantic content of the input text, thus laying a crucial, context-dependent structural foundation for accurate relational reasoning in the next step.

[0030] As shown in Figure 1, the method includes the following steps: Step S300, Relationship-Aware Graph Neural Network Reasoning. This step is another core component of the method, aiming to perform deep information aggregation and interaction on the dynamic multi-relation graph constructed in Step S200, generating event representations rich in complex relational semantics. Its goal is not only to enhance the representation of individual event nodes, but also to explicitly model and coordinate the inherent logic between different relation types, thereby achieving true joint reasoning. The specific implementation is as follows... Figure 4As shown, this invention comprises two layers: First, relation-specific graph attention information aggregation is performed. For each relation type corresponding to a dynamic graph, a multi-head graph attention network is used. This network calculates the attention weights between each event node and its neighboring nodes within the specific semantic space defined by the relation. These weights dynamically reflect the differences in importance of different neighboring information in the current relation context. Subsequently, the model performs a weighted summation of the features of neighboring nodes based on these weights, thereby aggregating local structural information onto the central node, achieving the first update of the node representation. Through this process, each event node obtains a set of enhanced representations, where each representation specifically encodes the semantic information of the node under a specific relation type. Second, a relation interaction layer is introduced for cross-relation joint reasoning. To overcome the limitations of processing different relation types in isolation, this invention designs a relation interaction layer. The input to this layer is the set of dedicated representations for each event node under different relations obtained in the preceding steps. Its core operation is to perform a cross-relation attention computation, allowing the representation of each relation to be dynamically adjusted and updated based on the representations of other relations. For example, the latest representation of a causal relationship references the representation of its temporal relationship, thus naturally learning logical constraints such as "cause precedes effect." Similarly, the representation of a coreference relationship is also influenced by its temporal or causal context to maintain consistency with other event relationships. Through this explicit interaction mechanism, the model can collaboratively reason about all relationships, effectively reducing logical conflicts in prediction results and significantly improving the accuracy and consistency of multi-label predictions in complex scenarios.

[0031] The output of this step is an event representation after deep relational reasoning. These representations deeply integrate the semantics of the event itself, its local graph structure information, and global relational interaction constraints, providing strong feature support for the final high-precision relational classification.

[0032] As shown in Figure 1, the method includes the following steps: Step S400, multi-label relationship classification and joint optimization. This step is the final output and global optimization stage of this method. It aims to accurately determine the multiple semantic relationships existing between any event mentions based on the event representation obtained in the preceding steps, which deeply integrates semantics, structure, and relational logic. Furthermore, it ensures the logical consistency of the overall prediction results by introducing prior knowledge constraints. Its specific implementation is as follows... Figure 5 As shown:

[0033] First, event pair representation construction and multi-label classification are performed. For any event pair requiring relation judgment, the enhanced representations processed in step S300 are combined to form a joint feature representation of the event pair. This joint representation is then input into a classifier module. This module sets independent classification channels for each predefined relation type. Each channel maps high-dimensional features to a scalar score through a fully connected layer, and then converts the score into an independent probability value between 0 and 1 through a sigmoid activation function. This probability value represents the confidence that the event pair has this specific relation. Since the calculation of each relation channel is independent, the model is allowed to predict that an event pair has multiple relations simultaneously, thus achieving true multi-label classification. Second, end-to-end joint training and optimization are implemented. The model is trained by minimizing a carefully designed joint loss function. This loss function is not a simple summation of task losses, but consists of two parts. The first part is the task fitting loss. A pointwise binary cross-entropy loss function is used to calculate the difference between the predicted probability and the true label for each relation category, ensuring the model's basic recognition ability for each relation. The second part is the logistic constraint loss. To improve the global logical consistency among predictions of different relationships, a constraint loss term based on prior knowledge is introduced. This constraint term, in a differentiable manner, transforms domain knowledge (e.g., "if there is a causal relationship between two events, then there must be a temporal relationship between them") into a mathematical model and penalizes combinations of predictions that violate such logical rules. For example, if the model predicts a causal relationship with high confidence but predicts a temporal relationship with low confidence, the constraint loss term will produce a large loss value, thereby guiding the model to correct this logical conflict during gradient descent.

[0034] The joint loss function is a weighted sum of the two losses mentioned above, which jointly optimizes all parameters of the model through end-to-end training. This design ensures that while the model accurately identifies various relationships, its output also conforms to the logical laws of the real world, significantly reducing self-contradictory phenomena in prediction results, thereby improving the overall reliability and interpretability of the system.

[0035] As a further improvement of the present invention, the preferred implementation of the operation of dynamically calculating the association weights between nodes in the dynamic multi-relationship graph construction process of step S200 is described in detail below. The core of this mechanism lies in designing an independent and learnable attention coefficient-based calculation module for each predefined relationship type (e.g., temporal, causal, coreference). This module aims to specifically evaluate and quantify the association strength or connection probability between any two event nodes under a specific relationship semantics. The specific calculation process includes the following steps: First, for the current relationship type to be calculated, the semantic representation vectors of the two event nodes are projected through a relationship-specific linear transformation layer. This projection operation aims to map the general semantic representation of the nodes to a dedicated feature space that is more consistent with the essence of the relationship. Second, the projected features of the two nodes are combined (e.g., by concatenation) and passed through a feedforward neural network layer. This neural network layer is usually composed of a linear transformation layer and a nonlinear activation function (such as LeakyReLU), and its function is to calculate the original association score of the two nodes under the context of the relationship. Subsequently, in order to obtain stable and comparable weights, the original score is normalized. Typically, the softmax function is used to normalize the raw scores of all neighbors of a node, resulting in a set of weight coefficients that satisfy the probability distribution property (i.e., the sum of all weights is 1). Ultimately, the normalized weight coefficients represent the strength or importance of information received by a node from its neighboring nodes from the perspective of a specific relationship, that is, the strength of the dynamic connection between nodes.

[0036] As a further improvement of the present invention, the dynamic multi-relationship graph construction process in step S200 includes an external knowledge enhancement step, which aims to integrate structured common sense and domain knowledge into the model's reasoning process to compensate for potential knowledge blind spots or ambiguities that may exist when learning solely from textual context, thereby significantly improving the accuracy and rationality of event relationship judgments. This step achieves the injection and fusion of external knowledge through a systematic process, the specific process of which is as follows: Figure 6As shown: First, external knowledge is queried and acquired. For event mention pairs with a relationship to be judged, their trigger words and core arguments are used as query keywords to retrieve information from one or more pre-built structured knowledge bases, such as the general commonsense base ConceptNet, domain knowledge bases such as UMLS in the biomedical field, or financial knowledge bases. The retrieval target is assertive knowledge related to the event pair that can corroborate or refute its specific relationship. Second, knowledge evidence is encoded and vectorized. The retrieval results returned from the knowledge base are normalized. The retrieval results typically include structured information such as relationship type, confidence score, and evidence path length. Subsequently, through a learnable embedding layer or lightweight encoder, this discrete evidence information is transformed into a fixed-length, numerical knowledge vector representation. This vector condenses the semantic information of supporting or disproving evidence related to the current event pair obtained from external sources. Finally, adaptive fusion based on a gating mechanism is performed. The external knowledge vector obtained above is concatenated or combined with the joint representation of the event pair generated internally by the model. This combined result is then input into a gating fusion network. The core function of this network is to dynamically generate a gating value between 0 and 1, which intelligently weighs the relative importance of external knowledge and internal evidence in the final decision. The calculation of the gating value depends on the current context and the confidence level of the knowledge itself. Finally, the final joint representation of the model is obtained by weighted summation of the external knowledge representation and the internal evidence representation, which are controlled by the gating value.

[0037] Through the aforementioned mechanism, the model can adaptively choose whether and to what extent to adopt external knowledge. For example, when there is relevant evidence with high confidence in the external knowledge base, the gate value approaches 1, and the model prioritizes relying on external knowledge; when there are no relevant records in the external knowledge base or the evidence quality is low, the gate value approaches 0, and the model mainly relies on internal reasoning results. This design enhances the rationality of reasoning and ensures the robustness of the model in the absence of external knowledge. This external knowledge enhancement step, by combining symbolic world knowledge with a data-driven deep learning model in a differentiable manner, effectively improves the model's understanding of common sense and domain logic, reduces erroneous predictions that contradict common sense or violate domain rules, and enhances the interpretability and reliability of the entire system.

[0038] Through this mechanism, the model can adaptively and accurately learn the unique connection patterns for each relation without relying on any predefined graph structure. This design greatly enhances the model's ability to distinguish between different relational semantics and its adaptability to complex textual contexts, thus providing a more reliable and accurate structured information foundation for subsequent relational reasoning.

[0039] As a further improvement of the present invention, in the relational reasoning process of step S300, the implementation of the relational interaction layer preferably adopts a mechanism based on cross-relational attention computation to explicitly model and fuse semantic dependencies and logical constraints between different relation types. Specifically, the operation process of the relational interaction layer is as follows: For each event node, its special representations under various relations such as temporal, causal, and coreference are combined into a relational representation set. This set is regarded as a sequence and a self-attention mechanism is applied for computation. In this process, query, key, and value vectors are generated for each relation. By calculating the similarity between query and key vectors between different relations, a relation-relational attention weight matrix is ​​obtained. This weight matrix dynamically quantifies the degree of mutual influence between any two relations in the current context. For example, the attention weight of causal relation representation to temporal relation representation will be significantly higher, which essentially encodes the common-sense logic that "causal events should precede result events" into parameters that the model can learn. Subsequently, the value vectors of each relation are weighted and fused using this attention weight matrix to generate a set of new relation-aware representations enhanced by interaction. After this operation, the representation of each relation incorporates the semantic information of other related relations and is no longer isolated.

[0040] The implementation of this cross-relational attention mechanism enables the model to automatically learn and reinforce inherent logic between relations, such as causal implicit timing and coreference requirements for relational consistency, and enforce these soft constraints during joint inference. This design significantly improves the model's inference consistency when dealing with complex scenarios involving multiple co-occurring relations, effectively reduces logical conflicts in prediction results, and thus enhances the accuracy and reliability of the entire system.

[0041] As a further improvement of this invention, the meta-learning adaptation in step S400 aims to significantly enhance the model's rapid adaptability and generalization performance when facing new domains, new relational patterns, or extremely scarce labeled data. The core of this meta-learning adaptation step is to enable the model to possess the ability to "learn how to learn," rather than merely memorizing specific patterns from the training data. Its specific implementation process is as follows: Figure 7As shown: First, task context awareness and task embedding generation. When the model needs to adapt to a new task support set containing a small number of labeled samples, it first performs forward propagation on all samples in the support set to obtain the high-order event pair representation of each sample output by the model's joint inference layer. Then, an aggregation strategy is used to synthesize the individual information of these samples into a fixed-length task embedding vector that can summarize the overall feature distribution of the current task. This vector encodes the key semantic context of the new task. Second, parameter increment prediction and rapid model fine-tuning. The generated task embedding vector is input into a dedicated, lightweight meta-learning adapter network, typically composed of one or more fully connected layers. This adapter network is meta-trained, and its function is to predict a set of targeted, small-amplitude parameter increments to the core parameters of the original model based on the input task embedding, rather than directly outputting new parameters. This adjustment reflects the specific transformation required from existing knowledge to the new task. Finally, adaptive model parameter updates. The predicted parameter increments are superimposed on the corresponding original parameters of the model, thereby achieving rapid and accurate adjustment of the model's behavior. This process is an efficient feedforward parameter update that avoids the overfitting problem that may result from performing a full backpropagation on a small amount of data, and achieves rapid task adaptation with a very small number of labeled samples.

[0042] By introducing this meta-learning adaptation step, the method described in this invention can effectively utilize the general event relationship knowledge learned on large source datasets. When faced with a new scenario with only a few dozen labeled samples, adaptation can be completed through a single or small amount of feedforward computation, greatly reducing the dependence on large-scale labeled data and enhancing the feasibility and practicality of the technology in real-world low-resource applications.

[0043] Secondly, the present invention provides an event relationship extraction system for implementing the event relationship extraction method as described in the first aspect. This system, through modular design, instantiates the process of the method into an organically integrated hardware and software architecture. Its core functional modules include: (1) an event encoding module, configured to receive raw text input and execute step S100. This module is specifically responsible for preprocessing the text, word segmentation, and event mention recognition, and calls a pre-trained deep language model to generate a contextual semantic embedding vector for each event mention, providing a high-quality semantic feature foundation for subsequent processing. (2) a dynamic graph construction module, connected to the event encoding module, configured to execute step S200. This module takes event mentions and their semantic vectors as input, dynamically calculates the association weight between any two event nodes for each predefined relationship type, and generates a series of relationship-specific adjacency matrices, thereby constructing multiple dynamic graph structures adaptive to the semantics of the input text. (3) a relationship reasoning module, connected to the dynamic graph construction module, configured to execute step S300. This module receives the dynamic graph structure and performs message passing and node representation updates in the semantic space of each relation through a relation-aware graph attention network. Its core is the integration of a relation interaction layer, which explicitly models and integrates the interaction and logical constraints between different relation representations through a cross-relation attention mechanism, and outputs an event-enhanced representation after deep joint reasoning. (4) Relation classification module, connected to the relation reasoning module, is configured to execute step S400. This module receives the enhanced event representation and, for any candidate event mention pair, independently judges the probability of its existence of various relations through a multi-label classifier. During training, this module optimizes a joint objective function that combines task fitting loss and logical constraint loss to achieve end-to-end parameter learning, ensuring that the output results have both high accuracy and logical consistency. Further, the system may also include a meta-learning adapter module. This module is connected to the relation reasoning module and / or relation classification module and is configured to run in low-resource scenarios. By analyzing a small number of support set samples, it generates task embedding vectors and predicts the incremental adjustments to the model's core parameters, thereby achieving rapid model adaptation and significantly improving its generalization ability and deployment efficiency when facing new domains or tasks. Furthermore, the system may also include an external knowledge access and fusion module. This module is configured to query common sense or domain evidence related to event pairs from an external structured knowledge base, encode it into a vectorized representation, and adaptively weight and integrate it with the model's internal representation through a gating fusion mechanism, thereby enhancing the rationality and robustness of the inference results. Through tight data coupling and collaborative work, the various modules of this system jointly achieve accurate, efficient, and reliable extraction of complex event relationships from text.

[0044] Thirdly, the present invention provides a computing device for event relation extraction. The device includes at least one memory and at least one processor; the memory stores executable instructions and model parameter data for event relation extraction; the processor implements the event relation extraction method as described in any one of the first aspects by reading and executing the executable instructions stored in the memory. Specifically, the memory is a non-volatile computer-readable storage medium, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), used to persistently store computer program code, pre-trained language model parameters, dynamic graph neural network weights, and meta-learning adapter parameters required to implement the method of the present invention. The processor is a central processing unit (CPU), graphics processing unit (GPU), tensor processing unit (TPU), or any combination thereof, configured to execute the instructions in the memory, specifically including: controlling the event encoding module to call the pre-trained model to encode input text; scheduling the dynamic graph construction module to calculate a relation-specific adjacency matrix; performing graph attention calculation and cross-relation interaction in the relation inference module; running the relation classification module to perform multi-label prediction and joint loss calculation; and, when enabled, calling the meta-learning adapter module or an external knowledge fusion module. The computing device receives input text data through a network interface, processes it through the processor, and outputs a set of structured event relationship triples (event mention 1, relationship type, event mention 2).

[0045] Fourthly, the present invention provides a computer-readable storage medium. This medium is a non-transitory computer-readable storage medium, such as an optical disc, USB flash drive, hard disk, server storage space, or distributed cloud storage space. It stores a computer program (or computer instructions, program code) that, when loaded and executed by a processor (such as the processor described in the third aspect), can automatically complete and implement all or part of the steps of the event relation extraction method as described in any one of the first aspects. The program code includes an instruction set for implementing event mention recognition and semantic encoding, an instruction set for implementing dynamic multi-relation graph construction, an instruction set for implementing relation-aware graph neural network reasoning, and an instruction set for implementing multi-label relation classification and joint optimization. Optionally, the program code further includes an instruction set for implementing meta-learning adaptation steps and an instruction set for introducing and fusing external knowledge.

[0046] In summary, this invention provides an efficient, robust, and highly interpretable solution for event relationship extraction by effectively combining dynamic graph construction, relational interaction reasoning, logical constraint loss, and meta-learning mechanisms. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and practicing the invention. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or conventional techniques in the art not disclosed herein.

Claims

1. A method for extracting event relationships based on dynamic graph relation reasoning, characterized in that, Includes the following steps: (1) Perform event mention recognition and semantic encoding on the input text to generate a contextual semantic representation vector of the event mention; (2) Based on the semantic representation vector, a multi-relationship event graph is dynamically constructed and common sense relationship evidence related to the event mention is queried from the structured knowledge base. External knowledge is adaptively integrated into the event representation through a gating mechanism to generate knowledge-enhanced event representation. (3) Based on the knowledge-enhanced event representation, a relation-aware graph neural network is used to perform message passing and node updates on the multi-relationship event graph to obtain a relation-enhanced relational interaction event representation; through explicit modeling across the relational interaction layer, the mutual influence and logical constraints between different relation types are modeled; the multi-relationship representations after interaction are fused to generate the final relation-enhanced event representation; (4) Based on the enhanced event representation, perform multi-label relationship classification on event mentions and predict the relationship type they exist; We set logical constraints to enhance the constraint loss term for logical consistency; we adopt end-to-end joint optimization to simultaneously optimize event mention recognition and relationship classification tasks; we introduce a meta-learning adaptation mechanism to encode the semantic features of the current task through task embedding vectors and generate model parameter increments based on task embeddings, so as to enable the model to adapt quickly in low-resource scenarios.

2. The "event mention identification and semantic encoding" method as described in claim 1, characterized in that: The text is encoded using a pre-trained language model to extract context-dependent word vector representations; event mentions in the text are identified through sequence labeling or span classification methods; for each event mention, its triggering words and contextual information are fused to generate an initial semantic representation vector.

3. The method for "dynamically constructing a multi-relationship event graph" as described in claim 1, characterized in that: The association weights between nodes are calculated independently for each relationship type, generating a relationship-specific adjacency matrix. The association weights are dynamically learned through an attention mechanism, reflecting the semantic association strength of event mentions under that relationship type.

4. The method of "adaptively integrating external knowledge into event representation through a gating mechanism" as described in claim 3, characterized in that: The system queries common-sense relationships related to event mentions from a structured knowledge base; external knowledge is incorporated into event representations through a gating mechanism to enhance the accuracy of relational reasoning.

5. The method for "using a relation-aware graph neural network to perform message passing and node updating on the multi-relation event graph" as described in claim 1, characterized in that: For each relation type, aggregate the representations of neighboring nodes in the relation-specific space; use gating mechanisms or attention weights to perform weighted fusion of neighboring information; and update the representation of the current node through nonlinear transformation.

6. The "explicit modeling across relational interaction layers" method as described in claim 1, characterized in that: By employing a cross-relational attention mechanism, we model the mutual influence and logical constraints between different relation types; we then fuse the multi-relational representations after interaction to generate the final relation-enhanced event representation.

7. The "multi-label relationship classification" method as described in claim 1, characterized in that: For each event mention pair, its relationship enhancement representation is concatenated; the probability of each relationship type is independently calculated using a fully connected layer and a sigmoid activation function; a threshold is set or a ranking strategy is adopted to determine the final set of predicted relationship labels.

8. The method for "setting logical constraints" as described in claim 1, characterized in that... The method also includes constraint loss terms designed based on prior logical rules between relationships, including at least one of the following: constraint loss of causal relationships implying temporal relationships; constraint loss of consistency of co-referenced events with third-party events; constraint loss of causal directionality; and constraint loss between mutually exclusive relationships.

9. The "end-to-end joint optimization" method as described in claim 1, characterized in that: A multi-task learning framework is adopted to simultaneously optimize event mention recognition and relation classification tasks; the loss function includes a binary cross-entropy loss for relation classification, as well as a constraint loss term to enhance logical consistency.

10. The "meta-learning adaptation mechanism" method as described in claim 1, characterized in that: The semantic features of the current task are encoded by task embedding vectors; model parameter increments are generated based on task embeddings to enable the model to adapt quickly in low-resource scenarios.

11. An event relationship extraction system, characterized in that, include: The event encoding module is used to perform step (1) of claim 1; the dynamic graph construction module is used to perform step (2) of claim 1; the relation-aware graph neural network inference module is used to perform step (3) of claim 1; and the multi-label relation classification and joint optimization module is used to perform step (4) of claim 1.

12. A computing device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-10.

Images