TDJEE: A Document-Level Joint Model for Financial Event Extraction

Abstract

Extracting financial events from numerous financial announcements is very important for investors to make right decisions. However, it is still challenging that event arguments always scatter in multiple sentences in a financial announcement, while most existing event extraction models only work in sentence-level scenarios. To address this problem, this paper proposes a relation-aware Transformer-based Document-level Joint Event Extraction model (TDJEE), which encodes relations between words into the context and leverages modified Transformer to capture document-level information to fill event arguments. Meanwhile, the absence of labeled data in financial domain could lead models be unstable in extraction results, which is known as the cold start problem. Furthermore, a Fonduer-based knowledge base combined with the distant supervision method is proposed to simplify the event labeling and provide high quality labeled training corpus for model training and evaluating. Experimental results on real-world Chinese financial announcement show that, compared with other models, TDJEE achieves competitive results and can effectively extract event arguments across multiple sentences.

DocEE: A Large-Scale and Fine-grained Benchmark for Document-level Event Extraction

“…WikiEvents and RAMS (Ebner et al, 2020) consist of 246/9,124 documents with only 59/65 argument types, and most of the arguments in the two datasets are shared among different event types without further refinement. Doc2EDAG, TDJEE and GIT (Zheng et al, 2019;Wang et al, 2021;Xu et al, 2021) only define 5 event types and 35 argument types in financial domain. Cancer Genetics, EPM, GENIA2011, GE-NIA2013, Pathway Curation and MLEE Ohta et al, 2011;Kim et al, 2011Kim et al, , 2013Van Landeghem et al, 2013) are limited to the biological domain.…”

“…RAMS (Ebner et al, 2020) limits the scope of the arguments in a 5-sentence window around its event trigger, which is not in line with the actual application, and the number of the argument types in RAMS is only 65, which is quite limited. Doc2EDAG, TDJEE and GIT (Zheng et al, 2019;Wang et al, 2021;Xu et al, 2021) contain only 5 event types and 35 argument types in financial domain. In summary, existing datasets for documentlevel EE fail in the following aspects: small scale of data, limited coverage of domain and insufficient refinement of argument types.…”

Learning from Limited Labels for Long Legal Dialogue

“…By contrast, much of the text in the criminal context serves the purpose of surfacing, discussing, and correcting case factors, which are not necessarily relational. This means parole hearings pose both extractive and abstractive tasks, often across multiple sentences, which is known to be challenging even in more structured settings (Wang et al, 2021).…”