A Joint Learning Framework for Document-Level Event Extraction

Bin Jiang, Sendong Zhu, Junyi Wu, Yu Liu

Abstract:

Document-level event extraction (DEE) is a challeng
ing task because arguments scattered across multiple sentences
form variable-length event lists. Existing methods often adopt
an autoregressive approach, performing event identification and
argument extraction sequentially. On one hand, this limits the
interaction between global event and local argument information;
on the other hand, errors in event type classification can
propagate to later argument extraction. To address these issues,
we propose an event- and argument-aware attention mechanism
to reduce error propagation, and a joint learning framework
(JLF) to enhance the interaction between event and argument
information. In addition, we design a complete event topology
decomposition (ETD) that supports the extraction of variable
length event lists across multiple sentences. Extensive experiments
show that our method achieves new state-of-the-art performance
on three public datasets, with improvements of 10.6% on the
ChFinAnn dataset, 5.6% on the DuEE-Fin dataset, and 14.8%
on the FNDEE dataset.