Dynamic Global Memory for Document-level Argument Extraction

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5264 - 5275 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Dynamic Global Memory for Document-level Argument Extraction Xinya Du Sha Li Heng Ji Department of Computer Science University of Illinois Urbana-Champaign {xinyadu2,shal2,hengji}@illinois.edu Abstract Extracting informative arguments of events from news articles is a challenging problem in information extraction, which requires a global contextual understanding of each document. While recent work on document-level extrac- tion has gone beyond single-sentence and in- creased the cross-sentence inference capability of end-to-end models, they are still restricted by certain input sequence length constraints and usually ignore the global context between events. To tackle this issue, we introduce a new global neural generation-based framework for document-level event argument extraction by constructing a document memory store to record the contextual event information and leveraging it to implicitly and explicitly help with decoding of arguments for later events. Empirical results show that our framework out- performs prior methods substantially and it is more robust to adversarially annotated exam- ples with our constrained decoding design.1 1 Introduction An event is a specific occurrence involving par- ticipants (people, objects, etc.). Understanding events in the text is necessary for building ma- chine reading systems, as well as for downstream tasks such as information retrieval, knowledge base population, and trend analysis of real-life world events (Sundheim, 1992). Event Extraction has long been studied as a local sentence-level task (Gr- ishman and Sundheim, 1996; Ji and Grishman, 2008b; Grishman, 2019; Lin et al., 2020). This has driven researchers to focus on developing ap- proaches for sentence-level predicate-argument ex- traction. This is problematic when events and their arguments spread across multiple sentences – in real-world cases, events are often written through- 1Our code and resources are available at https:// github.com/xinyadu/memory_docie for research purpose. … [S3] After having a shootout with several [policemen including Collin] last Thursday, both [Tamerlan Tsarnaev] and his younger brother [Dzhokhar] were captured a day later. … [S6] Two week ago, in Boston, authorities on Wednesday reopened [Boylston Street], the city thoroughfare where the explosion occurred near the finish line of the race. [S7] … a memorial service for campus policeman Sean Collin, who authorities say the brothers shot to death three days after the bombings … Trigger “captured” Jailer “policemen including Collin” Detainee “Tamerlan Tsarnaev”, “Dzhokhar” Event 1: Arrest Trigger “explosion” Attacker “Tamerlan Tsarnaev”, “Dzhokhar” Place “Boylston Street” Event 2: Attack-Detonate Figure 1: Document-level event argument extraction. out a document.2 In Figure 1, the excerpt of a news article de- scribes two events in the 3rd sentence (an arrest event triggered by “captured”) and the 6th sentence (an attack event triggered by “explosion”). S6 on its own contains little information about the ar- guments/participants of the explosion event, but together with the context of S3 and S7, we can find the informative arguments for the ATTACKER role. In this work, we focus on the informative argument extraction problem, which is more prac- …

Content truncated to 500 words