Extracting Temporal Relationships of Chinese Events Based on Graph Representation Learning

Abstract

Abstract The extraction of temporal relationships between events is crucial for downstream tasks such as automatic summarization and question-answering systems. Despite the impressive performance of large language models like ChatGPT in handling a wide range of tasks, extracting temporal relationships among Chinese events remains a significant challenge. Additionally, the lack of Chinese event temporal relation extraction datasets has constrained the development of this field. To address these issues, we have constructed MAND, a large-scale document-level dataset for evaluating event extraction and event temporal relation extraction. Subsequently, based on the MAND dataset, we propose a graph masking-repair method grounded in graph representation learning, which effectively captures global features and learns interdependencies among events. Furthermore, we represent coreference relationships in the document as a graph through coreference resolution and train our model using a graph masking mechanism. During the testing phase, we introduce an uncertainty calculation strategy to determine the prediction order of edges, aiming to reduce propagation errors. Experimental results demonstrate that our model outperforms previous methods in the field of Chinese event temporal relation extraction.