Search

CN-121997938-A - Event relation extraction method and system based on dynamic graph relation reasoning

CN121997938ACN 121997938 ACN121997938 ACN 121997938ACN-121997938-A

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 steps of carrying out event mention recognition and semantic coding on input texts to generate semantic representation vectors, dynamically constructing a multi-relation event diagram based on the vectors, learning time sequence, causal and co-index relation weights among events through a relation awareness attention mechanism, carrying out message transmission and node updating through a relation awareness graph neural network, carrying out explicit modeling of logical constraints among the relations through a cross-relation interaction layer, carrying out multi-label relation classification based on enhanced event representations to support multiple relations of the same event pair, introducing an external knowledge fusion mechanism to enhance reasoning rationality, carrying out end-to-end optimization by adopting a joint loss function containing logical constraint items, and realizing rapid migration of a low-resource scene by combining a meta learning adaptation mechanism. The method solves the problem that the conventional method is difficult to process the co-occurrence of multiple relations, long-distance dependence and low resource adaptation, and remarkably improves the accuracy, robustness and generalization capability of the event relation extraction.

Inventors

  • LIU WENXUAN
  • XU LIN
  • WANG YUHANG

Assignees

  • 四川工商学院
  • 刘文萱

Dates

Publication Date
20260508
Application Date
20260129

Claims (13)

  1. 1. The event relation extraction method based on dynamic graph relation reasoning is characterized by comprising the following steps of: (1) Carrying out event mention recognition and semantic coding on the input text to generate a context semantic representation vector of the event mention; (2) Dynamically constructing a multi-relation event diagram based on the semantic representation vector, inquiring common sense relation evidence related to the event mention pair from a structured knowledge base, and adaptively integrating external knowledge into the event representation through a gating mechanism to generate knowledge-enhanced event representation; (3) Based on the knowledge enhanced event representation, carrying out message transmission and node updating on the multi-relationship event graph by adopting a relationship-aware graph neural network to obtain a relationship enhanced relationship interaction event representation; (4) Based on the event representation of the relation enhancement, multi-label relation classification is carried out on the event mention pairs, the existing relation types are predicted, logic constraint is set, constraint loss items of logic consistency are enhanced, end-to-end joint optimization is adopted, event mention identification and relation classification tasks are optimized at the same time, a meta-learning adaptation mechanism is introduced, semantic features of a current task are encoded through task embedding vectors, model parameter increment is generated based on task embedding, and rapid adaptation of a model under a low-resource scene is realized.
  2. 2. The method of claim 1, wherein the text is encoded using a pre-trained language model, a context-dependent word vector representation is extracted, event references in the text are identified by a sequence labeling or span classification method, and trigger words and context information are fused for each event reference to generate an initial semantic representation vector.
  3. 3. The method of claim 1, wherein the association weights among the nodes are calculated independently for each relationship type to generate a relationship-specific adjacency matrix, and the association weights are obtained through dynamic learning of an attention mechanism and reflect the semantic association strength of the event under the relationship type.
  4. 4. The method for adaptively incorporating external knowledge into an event representation through a gating mechanism as set forth in claim 3, wherein the query of the structured knowledge base for common sense relationships related to the event reference pairs and the incorporation of external knowledge into the event representation through the gating mechanism enhance the accuracy of relationship reasoning.
  5. 5. The method for message passing and node updating of the multiple relationship event graph by using a relationship-aware graph neural network according to claim 1, wherein for each relationship type, the representation of the neighbor node under the relationship specific space is aggregated, the neighbor information is weighted and fused by using a gating mechanism or attention weight, and the representation of the current node is updated by nonlinear transformation.
  6. 6. The method of claim 1, wherein interactions and logical constraints between different relationship types are modeled by a cross-relationship attention mechanism, and the interacted multi-relationship representations are fused to generate a final relationship enhancement event representation.
  7. 7. The method of claim 1, wherein each event is referred to as a pair, the relationship enhancement representation is spliced, the probability of each relationship type is calculated independently through a full connection layer and a sigmoid activation function, and a threshold is set or a ranking strategy is adopted to determine the final predicted relationship tag set.
  8. 8. The method of setting logical constraints of claim 1, further comprising designing constraint loss terms based on a priori logic rules between relationships, including at least one of constraint loss of causal relationship implicit timing relationships, constraint loss of co-referencing events on third party event relationship consistency, constraint loss of causal relationship directionality, constraint loss between mutual exclusion relationships.
  9. 9. The end-to-joint optimization method of claim 1, wherein a multitasking learning framework is employed to simultaneously optimize event-mentioning recognition and relationship classification tasks, and wherein the penalty function comprises a binary cross entropy penalty for relationship classification, and a constraint penalty term for enhancing logical consistency.
  10. 10. The method of claim 1, wherein the semantic features of the current task are encoded by task embedding vectors, and model parameter increments are generated based on task embedding, so that the model is quickly adapted in a low-resource scene.
  11. 11. An event relation extraction system is characterized by comprising an event coding module, a dynamic diagram construction module, a relation-aware graph neural network reasoning module and a multi-label relation classification and joint optimization module, wherein the event coding module is used for executing the step (1) of the claim 1, the dynamic diagram construction module is used for executing the step (2) of the claim 1, the relation-aware graph neural network reasoning module is used for executing the step (3) of the claim 1, and the multi-label relation classification and joint optimization module is used for executing the step (4) of the claim 1.
  12. 12. A computing device comprising a memory, a processor and a computer program stored on the memory, wherein the processor implements the method of any of claims 1-10 when executing the program.
  13. 13. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any of claims 1-10.

Description

Event relation extraction method and system based on dynamic graph relation reasoning Technical Field The invention belongs to the technical field of natural language processing and artificial intelligence, and particularly relates to a method and a system for extracting semantic relations between events from unstructured texts. In particular to an event relation extraction technology which dynamically constructs a multi-relation diagram and performs joint reasoning by utilizing a relation-aware diagram neural network. The method can be widely applied to scenes needing to deeply understand event logic in the text, such as public opinion analysis, intelligent questions and answers, knowledge graph construction and the like. Background Event relationship extraction is a core task in the field of information extraction, and aims to identify semantic relationships between event references in text, such as time sequence relationships, causal relationships, co-index relationships and the like. Accurate extraction of these relationships is critical to building an event knowledge chain, understanding narrative logic. The prior art scheme mainly faces the challenge of relation isolation processing that the traditional method often treats different types of relations (such as time sequence and cause and effect) as independent tasks, ignores inherent logic association and constraint between the relations, and for example, the cause and effect relation usually implies time sequence relations. Static modeling limitations-methods based on predefined rules or static graph structures (such as dependency syntax trees) lack self-adapting capability to context, and it is difficult to handle dynamically changing event associations in complex contexts. Long-distance dependency and sparsity when event references are scattered in long documents or marked data are scarce, the model has difficulty in capturing long-distance dependency relationships effectively and is prone to reduction in generalization capability due to insufficient data. The multi-relation co-occurrence conflict is that a pair of events always exist in a real text and have multiple relations, and the existing method has the defect of ensuring the logical consistency of multi-label prediction. Although some researches have been attempted to partially solve the above problems by joint learning, graph neural network or introducing external knowledge, there is still a breakthrough in how to dynamically and explicitly unify the interactions and constraints between various relations and effectively adapt to complex scenarios such as low resources, long documents, etc. Therefore, a new method for extracting event relationships is needed, which can uniformly model various relationship interactions, adaptively and dynamically construct semantic association, and has a strong generalization capability. The invention is based on a dynamic graph neural network and a meta learning mechanism, aims to solve the problems and provides an effective solution for event relation extraction under a complex context. Disclosure of Invention In order to achieve the aim of the invention, the invention adopts the following technical scheme that in the first aspect, the invention provides an event extraction method based on multi-scale semantic understanding and potential attention regulation, and the core technical scheme comprises four key modules, namely event mention identification and semantic coding, dynamic multi-relation graph construction, relation-aware graph neural network reasoning, multi-label relation classification and joint optimization. As a further improvement of the invention, the event mention recognition and semantic coding specifically comprises preprocessing input text and event mention recognition, and acquiring context-related semantic representation vectors of each event mention by utilizing a pre-trained language model. This step provides a high quality semantic basis for subsequent analysis. As a further improvement of the invention, the dynamic multi-relation graph construction specifically comprises the steps of dynamically constructing a multi-relation event graph, inquiring common sense relation evidence related to the event reference pair from a structured knowledge base, and adaptively integrating external knowledge into the event representation through a gating mechanism to generate a knowledge-enhanced event representation. As a further improvement of the invention, the relationship-aware graph neural network reasoning specifically comprises the steps of adopting a multi-head graph attention mechanism for a constructed dynamic multi-relationship graph, carrying out message transmission and node representation updating under a specific semantic space of each relationship, and generating an event representation with enhanced relationship. Further, the joint reasoning capability is improved through the mutual influence and logic constraint among different rel