Neural Event Extraction from Scientific Discourse

Abstract

Event Extraction (EE) plays a crucial role in Natural Language Processing, aiming to extract informative structures related to events. Events typically consist of a word that denotes the event (trigger) and additional event-specific information (arguments). High-quality structured information derived from Event Extraction can prove valuable for various downstream tasks, including Question Answering and Fact-Checking. In our research, we have delved into both intra-sentence (or sentence-level) EE, where the trigger and arguments are confined to a single sentence, and inter-sentence (or document-level) EE, where the scope extends beyond sentence boundaries. While intra-sentence EE has traditionally been the primary focus, the rise in computing power and the advent of Pre-trained Language Models (PLMs) have facilitated the growing interest in inter-sentence EE. Initially, our objectives revolve around addressing intra-sentence EE within the biomedical domain. Biomedical event extraction presents unique challenges, featuring nested events where the events themselves can serve as arguments for other events. We explore and compare methods for constructing event structures in a pipeline setting, emphasising computational efficiency without sacrificing significant performance. Subsequently, we explore the utilisation of PLMs to enhance the exploitation of contextual information. Contextual cues are integral to EE as they provide valuable insights into understanding events. Our investigation focuses on using markers -special tokens inserted into the input- to augment the language model with additional information. We propose both heuristic and automatic methods to identify relevant contextual cues and validate their performance through experiments. Lastly, we build upon the previous approach to address a demanding inter-sentence biomedical EE scenario. Utilising typed markers -markers of different types- we harness entity information to aid trigger identification and the construction of inter- sentence events. We demonstrate the effectiveness of our approach on a newly developed biomedical EE corpus that we introduce.

Date of Award	31 Dec 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Sophia Ananiadou (Supervisor) & Junichi Tsujii (Supervisor)