Detecting event location is a key aspect of event extraction from news and social media. However, this task has not received strong attention recently in comparison to event classification or identifying the event time and the semantic arguments of the event, such as victims, perpetrators, means of action, affected infrastructure, etc. Nevertheless, the location as an event argument plays a crucial role in all event detection applications: conflict detection, health threat monitoring, disaster impact assessment, etc. In addition recent advances in Natural Language Processing open new perspectives. This paper looks at the location detection as a text classification problem using Large Language Models. Our method relies on using a BERT model in the context of event extraction from news articles for classifying each location in the text into one of the two classes: place of an event, reported by the article, called main location, or another location mention, called here secondary location . Evaluation shows promising results on a news corpus of protest events and demonstrates the feasibility of our approach and the event geolocation task in general.

TANEV Hristo;  DE LONGUEVILLE Bertrand; 

2024-10-30

INCOMA Ltd., Shoumen, Bulgaria

JRC133176

1611-3349 (online),   

https://aclanthology.org/2023.case-1.2,    https://publications.jrc.ec.europa.eu/repository/handle/JRC133176,